AI search engine and MCP server for bots and agents. One Rust binary indexes code, docs, logs, JSONL/CSV and PDFs for BM25 full-text, vector and hybrid search, code search without grep, agent memory and RAG. Elasticsearch-compatible API; Corpus Hub serves hundreds of reference memory corpora.
Rust agent-memoryai-agentsbm25claude-codeclicode-searchcursordatabaseelasticsearchfull-text-searchhybrid-searchjsonllocal-firstmcpmcp-servermodel-context-protocolragrustsearch-enginevector-search
blog: one cover series for every post (field-log posters, og + hero + thumbnail) (#1918)
xerj-orgd7cf8a100025
diff --git a/landing/blog.css b/landing/blog.cssindex e717c2339..f6f9e2e4a 100644--- a/landing/blog.css+++ b/landing/blog.css@@ -149,6 +149,16 @@.verdict-panel p strong { color:var(--z-ink); }.verdict-panel p code.inline { color:var(--z-ink); }+ /* cover art — the field-log poster series, rendered by+ scripts/seo/blog_covers/covers.py. The poster is a self-contained print+ on the brand's light paper, ink and one earned yellow (it doubles as the+ og card), so on either theme it sits in a flat mount: a 1px rule in the+ page's own line colour, square corners, no shadow. Only image rules live+ here; page colours belong to the theme. */+ .blog-cover { margin: 0 0 var(--sp-6); padding: 0; border: 1px solid var(--z-line);+ overflow: hidden; background: #f6f4ee; line-height: 0; }+ .blog-cover img { display: block; width: 100%; height: auto; aspect-ratio: 1600 / 640; }+/* ---- index components: bl-card, blp preview bars ---- *//* page-local — blog index: dated cards, preview visual left, brand tokensonly, 1px lines, no icons, no shadows, no filled shapes. */@@ -167,6 +177,10 @@}.bl-card:hover, .bl-card:focus-visible { border-color: var(--z-accent); }+ /* card thumbnail: the post's cover at 640x336 (og ratio), lazy-loaded */+ .bl-thumb { display: block; width: 100%; height: auto; aspect-ratio: 640 / 336; align-self: start;+ border: 1px solid var(--z-line); background: #f6f4ee; }+/* preview slot — a theme-native miniature scorecard, pure divs; no imagefile, correct in day and night by construction (site precedent: everychart in the benchmarks pages is CSS bars). */@@ -220,3 +234,52 @@.bl-card { grid-template-columns: minmax(0, 1fr); gap: var(--sp-4); }.bl-title { font-size: var(--fs-20); }}++/* ---- reading theme: light paper, AAA text (2026-10-20) ----+ Every /blog/ page is served in the day theme (data-theme="day" on <html>),+ so the light covers sit on the page colour they were printed for. Inside the+ blog the day tokens are tightened to WCAG AAA on paper #f6f4ee: body ink+ 17.9:1, secondary text 8.3:1, accent text 7.7:1. Brand ochre #A06800+ (4.27:1, too light for text) is kept for rules and borders. Flat: no+ gradients, no shadows. The NIGHT toggle still works (night tokens untouched). */+ html[data-theme='day'] {+ --z-ink: #0b0b0d;+ --z-mute: #4a4843;+ --z-accent: #6b4500;+ --z-ochre: #a06800;+ --z-cmp: #4f6a88;+ }+ html[data-theme='day'] pre.code { border-left-color: var(--z-ochre); }+ html[data-theme='day'] .bl-card:hover,+ html[data-theme='day'] .bl-card:focus-visible { border-color: var(--z-ochre); }++ /* type: size, line height and measure only */+ .uc-hero .uc-eyebrow, .blog-crumb, .blog-byline { font-size: 13px; }+ .uc-hero .lede { font-size: 21px; line-height: 1.6; max-width: 68ch; margin: var(--sp-8) 0 0; }+ section.scene-section > .kicker { font-size: 13px; margin-bottom: var(--sp-4); }+ section.scene-section > h2.scene { margin: 0 0 var(--sp-8); }+ section.scene-section > p,+ section.scene-section > ul > li,+ section.scene-section > ol > li { font-size: 19px; line-height: 1.7; }+ section.scene-section > p { max-width: 68ch; margin: 0 0 1.1em; }+ section.scene-section > p:last-child { margin-bottom: 0; }+ section.scene-section > ul,+ section.scene-section > ol { max-width: 68ch; margin: 0 0 1.1em; padding-left: 1.25em; }+ section.scene-section > ul > li,+ section.scene-section > ol > li { margin: 0 0 .6em; }+ section.scene-section > pre.code { font-size: 15px; line-height: 1.6;+ margin: var(--sp-6) 0 var(--sp-8); padding: 20px 22px; }+ section.scene-section > .vt-wrap { margin: var(--sp-6) 0 var(--sp-8); }+ section.scene-section > .stat-band { margin: var(--sp-6) 0 var(--sp-8); }+ table.vt { font-size: 15.5px; line-height: 1.5; }+ .vt th { font-size: 12.5px; }+ .stat .k { font-size: 12.5px; }+ .stat .s { font-size: 15px; }+ .bl-excerpt { font-size: 17px; line-height: 1.65; }+ @media (max-width: 620px) {+ .uc-hero .lede { font-size: 19px; }+ section.scene-section > p,+ section.scene-section > ul > li,+ section.scene-section > ol > li { font-size: 18px; }+ section.scene-section > pre.code { font-size: 13px; padding: 16px; }+ }diff --git a/landing/blog/covers/cutting-the-index-28-percent-hero.webp b/landing/blog/covers/cutting-the-index-28-percent-hero.webpnew file mode 100644index 000000000..a729087b8Binary files /dev/null and b/landing/blog/covers/cutting-the-index-28-percent-hero.webp differdiff --git a/landing/blog/covers/cutting-the-index-28-percent-thumb.webp b/landing/blog/covers/cutting-the-index-28-percent-thumb.webpnew file mode 100644index 000000000..1b988537eBinary files /dev/null and b/landing/blog/covers/cutting-the-index-28-percent-thumb.webp differdiff --git a/landing/blog/covers/does-xerj-beat-jev-hero.webp b/landing/blog/covers/does-xerj-beat-jev-hero.webpnew file mode 100644index 000000000..a5b0a6d62Binary files /dev/null and b/landing/blog/covers/does-xerj-beat-jev-hero.webp differdiff --git a/landing/blog/covers/does-xerj-beat-jev-thumb.webp b/landing/blog/covers/does-xerj-beat-jev-thumb.webpnew file mode 100644index 000000000..7eaf89688Binary files /dev/null and b/landing/blog/covers/does-xerj-beat-jev-thumb.webp differdiff --git a/landing/blog/covers/eval-scars-from-agent-memory-hero.webp b/landing/blog/covers/eval-scars-from-agent-memory-hero.webpnew file mode 100644index 000000000..16c14c15cBinary files /dev/null and b/landing/blog/covers/eval-scars-from-agent-memory-hero.webp differdiff --git a/landing/blog/covers/eval-scars-from-agent-memory-thumb.webp b/landing/blog/covers/eval-scars-from-agent-memory-thumb.webpnew file mode 100644index 000000000..761cbc270Binary files /dev/null and b/landing/blog/covers/eval-scars-from-agent-memory-thumb.webp differdiff --git a/landing/blog/covers/jev-vs-a-bm25-vote-hero.webp b/landing/blog/covers/jev-vs-a-bm25-vote-hero.webpnew file mode 100644index 000000000..dcd7e2130Binary files /dev/null and b/landing/blog/covers/jev-vs-a-bm25-vote-hero.webp differdiff --git a/landing/blog/covers/jev-vs-a-bm25-vote-thumb.webp b/landing/blog/covers/jev-vs-a-bm25-vote-thumb.webpnew file mode 100644index 000000000..e17797517Binary files /dev/null and b/landing/blog/covers/jev-vs-a-bm25-vote-thumb.webp differdiff --git a/landing/blog/covers/one-vulnerability-three-names-hero.webp b/landing/blog/covers/one-vulnerability-three-names-hero.webpnew file mode 100644index 000000000..231e31ee8Binary files /dev/null and b/landing/blog/covers/one-vulnerability-three-names-hero.webp differdiff --git a/landing/blog/covers/one-vulnerability-three-names-thumb.webp b/landing/blog/covers/one-vulnerability-three-names-thumb.webpnew file mode 100644index 000000000..b5957460aBinary files /dev/null and b/landing/blog/covers/one-vulnerability-three-names-thumb.webp differdiff --git a/landing/blog/covers/the-corpus-hub-hero.webp b/landing/blog/covers/the-corpus-hub-hero.webpnew file mode 100644index 000000000..7a330b30eBinary files /dev/null and b/landing/blog/covers/the-corpus-hub-hero.webp differdiff --git a/landing/blog/covers/the-corpus-hub-thumb.webp b/landing/blog/covers/the-corpus-hub-thumb.webpnew file mode 100644index 000000000..c9d189fbaBinary files /dev/null and b/landing/blog/covers/the-corpus-hub-thumb.webp differdiff --git a/landing/blog/covers/the-exploit-hub-hero.webp b/landing/blog/covers/the-exploit-hub-hero.webpnew file mode 100644index 000000000..56a5e74ebBinary files /dev/null and b/landing/blog/covers/the-exploit-hub-hero.webp differdiff --git a/landing/blog/covers/the-exploit-hub-thumb.webp b/landing/blog/covers/the-exploit-hub-thumb.webpnew file mode 100644index 000000000..8b6924a90Binary files /dev/null and b/landing/blog/covers/the-exploit-hub-thumb.webp differdiff --git a/landing/blog/covers/the-rc80-gates-hero.webp b/landing/blog/covers/the-rc80-gates-hero.webpnew file mode 100644index 000000000..da96ebee4Binary files /dev/null and b/landing/blog/covers/the-rc80-gates-hero.webp differdiff --git a/landing/blog/covers/the-rc80-gates-thumb.webp b/landing/blog/covers/the-rc80-gates-thumb.webpnew file mode 100644index 000000000..00c7550f8Binary files /dev/null and b/landing/blog/covers/the-rc80-gates-thumb.webp differdiff --git a/landing/blog/cutting-the-index-28-percent.html b/landing/blog/cutting-the-index-28-percent.htmlindex a14a5bddb..38e1423d7 100644--- a/landing/blog/cutting-the-index-28-percent.html+++ b/landing/blog/cutting-the-index-28-percent.html@@ -1,5 +1,5 @@<!DOCTYPE html>-<html lang="en">+<html lang="en" data-theme="day"><head><meta charset="UTF-8"><meta name="viewport" content="width=device-width, initial-scale=1.0">@@ -9,7 +9,7 @@<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin><link href="https://fonts.googleapis.com/css2?family=Big+Shoulders+Display:wght@400;700;800;900&family=Inter:wght@400;500;600;700&family=IBM+Plex+Sans:wght@400;500;600&family=JetBrains+Mono:wght@400;500;700&display=swap" rel="stylesheet"><link rel="stylesheet" href="/style.css?v=if9">-<link rel="stylesheet" href="/blog.css?v=1">+<link rel="stylesheet" href="/blog.css?v=4"><!-- seo:begin — generated by scripts/seo/fix_heads.py · do not hand-edit --><link rel="canonical" href="https://xerj.org/blog/cutting-the-index-28-percent"><meta property="og:site_name" content="XERJ.ai">@@ -17,15 +17,15 @@<meta property="og:title" content="Cutting the index 28%: dissect first, write Rust second"><meta property="og:description" content="Four file-format changes took our 100k-doc index from 5.52 MB to 3.99 MB (−28%), each gated on dissecting real segments first. The losing rows are in the post."><meta property="og:url" content="https://xerj.org/blog/cutting-the-index-28-percent">-<meta property="og:image" content="https://xerj.org/og/xerj-card.png">+<meta property="og:image" content="https://xerj.org/og/blog-cutting-the-index-28-percent.png"><meta property="og:image:width" content="1200"><meta property="og:image:height" content="630">-<meta property="og:image:alt" content="XERJ.ai — the search engine built for AI agents">+<meta property="og:image:alt" content="Cover: “Cutting the index 28%” in display type beside the post’s square-matrix art, labelled with its figures. The art: a large ink block of squares wrapped in a yellow L-shaped band that marks the 28% cut, beside two pairs of thin before-and-after bars"><meta property="og:locale" content="en_US"><meta name="twitter:card" content="summary_large_image"><meta name="twitter:title" content="Cutting the index 28%: dissect first, write Rust second"><meta name="twitter:description" content="Four file-format changes took our 100k-doc index from 5.52 MB to 3.99 MB (−28%), each gated on dissecting real segments first. The losing rows are in the post.">-<meta name="twitter:image" content="https://xerj.org/og/xerj-card.png">+<meta name="twitter:image" content="https://xerj.org/og/blog-cutting-the-index-28-percent.png"><script type="application/ld+json">{"@context": "https://schema.org",@@ -60,7 +60,7 @@"headline": "Cutting the index 28%: dissect first, write Rust second","description": "Four file-format changes took our 100k-doc index from 5.52 MB to 3.99 MB (−28%), each gated on dissecting real segments first. The losing rows are in the post.","url": "https://xerj.org/blog/cutting-the-index-28-percent",- "image": "https://xerj.org/og/xerj-card.png",+ "image": "https://xerj.org/og/blog-cutting-the-index-28-percent.png","inLanguage": "en","datePublished": "2026-09-27","dateModified": "2026-09-27",@@ -130,6 +130,7 @@<main><div class="uc-hero"><p class="blog-crumb"><a href="/blog/">← ALL POSTS · /BLOG</a></p>+ <figure class="blog-cover"><img src="/blog/covers/cutting-the-index-28-percent-hero.webp" width="1600" height="640" alt="Cover art: a large ink block of squares wrapped in a yellow L-shaped band that marks the 28% cut, beside two pairs of thin before-and-after bars" fetchpriority="high" decoding="async"></figure><div class="uc-eyebrow">STORAGE · ON-DISK SIZE · 100K-DOC HARNESS · RUNS 2026-09-26/27</div><h1>A DISSECTION, FOUR FILE FORMATS,<br><span class="accent">AND 28% OFF THE INDEX</span></h1><p class="blog-byline">XERJ ENGINEERING · PUBLISHED 2026-09-27</p>diff --git a/landing/blog/does-xerj-beat-jev.html b/landing/blog/does-xerj-beat-jev.htmlindex 19e927f00..22c5576eb 100644--- a/landing/blog/does-xerj-beat-jev.html+++ b/landing/blog/does-xerj-beat-jev.html@@ -10,7 +10,7 @@<link href="https://fonts.googleapis.com/css2?family=Big+Shoulders+Display:wght@400;700;800;900&family=Inter:wght@400;500;600;700&family=IBM+Plex+Sans:wght@400;500;600;700&family=JetBrains+Mono:wght@400;500;700&display=swap" rel="stylesheet"><link rel="icon" href="data:image/svg+xml,<svg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 100 100'><rect width='100' height='100' fill='%230b0b0d'/><text x='50' y='68' font-size='60' font-family='Big Shoulders Display,sans-serif' font-weight='900' fill='%23ffc400' text-anchor='middle'>X</text></svg>"><link rel="stylesheet" href="/style.css?v=if9">-<link rel="stylesheet" href="/blog.css?v=1">+<link rel="stylesheet" href="/blog.css?v=4"><!-- seo:begin — generated by scripts/seo/fix_heads.py · do not hand-edit --><link rel="canonical" href="https://xerj.org/blog/does-xerj-beat-jev"><meta property="og:site_name" content="XERJ.ai">@@ -18,15 +18,15 @@<meta property="og:title" content="Does XERJ beat JEV? On the bill, outright. On FiQA, no."><meta property="og:description" content="1,271 judged queries: Jev rerank takes SciFact 0.7410 and FiQA 0.3638 (+0.126), our zero-API-token hybrid keeps NFCorpus 0.3448, and the FiQA run cost $0.6375."><meta property="og:url" content="https://xerj.org/blog/does-xerj-beat-jev">-<meta property="og:image" content="https://xerj.org/og/does-xerj-beat-jev.png">+<meta property="og:image" content="https://xerj.org/og/blog-does-xerj-beat-jev.png"><meta property="og:image:width" content="1200"><meta property="og:image:height" content="630">-<meta property="og:image:alt" content="XERJ vs JEV, measured: SciFact, NFCorpus and FiQA nDCG@10 over 1,271 judged queries, and the FiQA judge bill — $0.6375 against $0 local">+<meta property="og:image:alt" content="Cover: “Does XERJ beat JEV?” in display type beside the post’s square-matrix art, labelled with its figures. The art: three pairs of square-matrix bars, ink for XERJ and blue for BM25+Jev, over one long blue bar and a single yellow square for the zero-cost local run"><meta property="og:locale" content="en_US"><meta name="twitter:card" content="summary_large_image"><meta name="twitter:title" content="Does XERJ beat JEV? On the bill, outright. On FiQA, no."><meta name="twitter:description" content="1,271 judged queries: Jev rerank takes SciFact 0.7410 and FiQA 0.3638 (+0.126), our zero-API-token hybrid keeps NFCorpus 0.3448, and the FiQA run cost $0.6375.">-<meta name="twitter:image" content="https://xerj.org/og/does-xerj-beat-jev.png">+<meta name="twitter:image" content="https://xerj.org/og/blog-does-xerj-beat-jev.png"><script type="application/ld+json">{"@context": "https://schema.org",@@ -61,7 +61,7 @@"headline": "Does XERJ beat JEV? On the bill, outright. On FiQA, no.","description": "1,271 judged queries: Jev rerank takes SciFact 0.7410 and FiQA 0.3638 (+0.126), our zero-API-token hybrid keeps NFCorpus 0.3448, and the FiQA run cost $0.6375.","url": "https://xerj.org/blog/does-xerj-beat-jev",- "image": "https://xerj.org/og/does-xerj-beat-jev.png",+ "image": "https://xerj.org/og/blog-does-xerj-beat-jev.png","inLanguage": "en","datePublished": "2026-09-21","dateModified": "2026-09-27",@@ -132,6 +132,7 @@<main><div class="uc-hero"><p class="blog-crumb"><a href="/blog/">← ALL POSTS · /BLOG</a></p>+ <figure class="blog-cover"><img src="/blog/covers/does-xerj-beat-jev-hero.webp" width="1600" height="640" alt="Cover art: three pairs of square-matrix bars, ink for XERJ and blue for BM25+Jev, over one long blue bar and a single yellow square for the zero-cost local run" fetchpriority="high" decoding="async"></figure><div class="uc-eyebrow">BENCHMARKS · ONE JUDGE MODEL, MEASURED · ONE WIRE, ANSWERED · RUNS 2026-09-19..21</div><h1>DOES XERJ BEAT JEV?<br><span class="accent">ON THE BILL, OUTRIGHT. ON FIQA, NO.</span></h1><p class="blog-byline">AUTHOR · XERJ ENGINEERING TEAM · PUBLISHED · 2026-09-21 · UPDATED · 2026-09-27</p>diff --git a/landing/blog/eval-scars-from-agent-memory.html b/landing/blog/eval-scars-from-agent-memory.htmlindex d6ad2d650..fb6905056 100644--- a/landing/blog/eval-scars-from-agent-memory.html+++ b/landing/blog/eval-scars-from-agent-memory.html@@ -10,7 +10,7 @@<link href="https://fonts.googleapis.com/css2?family=Big+Shoulders+Display:wght@400;700;800;900&family=Inter:wght@400;500;600;700&family=IBM+Plex+Sans:wght@400;500;600;700&family=JetBrains+Mono:wght@400;500;700&display=swap" rel="stylesheet"><link rel="icon" href="data:image/svg+xml,<svg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 100 100'><rect width='100' height='100' fill='%230b0b0d'/><text x='50' y='68' font-size='60' font-family='Big Shoulders Display,sans-serif' font-weight='900' fill='%23ffc400' text-anchor='middle'>X</text></svg>"><link rel="stylesheet" href="/style.css?v=if9">-<link rel="stylesheet" href="/blog.css?v=1">+<link rel="stylesheet" href="/blog.css?v=4"><!-- seo:begin — generated by scripts/seo/fix_heads.py · do not hand-edit --><link rel="canonical" href="https://xerj.org/blog/eval-scars-from-agent-memory"><meta property="og:site_name" content="XERJ.ai">@@ -18,15 +18,15 @@<meta property="og:title" content="Six eval scars from a production agent-memory system"><meta property="og:description" content="Guest recipes from Chunxiao Wang (Assay): topic slicing beat token windows on recall@1 by double digits, a 0.87 classifier nearly matched by a 0.78 always-yes baseline, and the mask bug that scored every query zero. Plus the corpus gap we checked and confirmed."><meta property="og:url" content="https://xerj.org/blog/eval-scars-from-agent-memory">-<meta property="og:image" content="https://xerj.org/og/xerj-card.png">+<meta property="og:image" content="https://xerj.org/og/blog-eval-scars-from-agent-memory.png"><meta property="og:image:width" content="1200"><meta property="og:image:height" content="630">-<meta property="og:image:alt" content="XERJ.ai — the search engine built for AI agents">+<meta property="og:image:alt" content="Cover: “Six eval scars from production agent memory” in display type beside the post’s square-matrix art, labelled with its figures. The art: six clusters of diagonal scratch marks made of small squares, five in ink and one in yellow"><meta property="og:locale" content="en_US"><meta name="twitter:card" content="summary_large_image"><meta name="twitter:title" content="Six eval scars from a production agent-memory system"><meta name="twitter:description" content="Guest recipes from Chunxiao Wang (Assay): topic slicing beat token windows on recall@1 by double digits, a 0.87 classifier nearly matched by a 0.78 always-yes baseline, and the mask bug that scored every query zero. Plus the corpus gap we checked and confirmed.">-<meta name="twitter:image" content="https://xerj.org/og/xerj-card.png">+<meta name="twitter:image" content="https://xerj.org/og/blog-eval-scars-from-agent-memory.png"><script type="application/ld+json">{"@context": "https://schema.org",@@ -61,7 +61,7 @@"headline": "Six eval scars from a production agent-memory system","description": "Guest recipes from Chunxiao Wang (Assay): topic slicing beat token windows on recall@1 by double digits, a 0.87 classifier nearly matched by a 0.78 always-yes baseline, and the mask bug that scored every query zero. Plus the corpus gap we checked and confirmed.","url": "https://xerj.org/blog/eval-scars-from-agent-memory",- "image": "https://xerj.org/og/xerj-card.png",+ "image": "https://xerj.org/og/blog-eval-scars-from-agent-memory.png","inLanguage": "en","datePublished": "2026-10-07","dateModified": "2026-10-11",@@ -132,6 +132,7 @@<main><div class="uc-hero"><p class="blog-crumb"><a href="/blog/">← ALL POSTS · /BLOG</a></p>+ <figure class="blog-cover"><img src="/blog/covers/eval-scars-from-agent-memory-hero.webp" width="1600" height="640" alt="Cover art: six clusters of diagonal scratch marks made of small squares, five in ink and one in yellow" fetchpriority="high" decoding="async"></figure><div class="uc-eyebrow">GUEST CONTRIBUTION · AGENT MEMORY · RETRIEVAL EVAL · ISSUES #1118 + #1138</div><h1>SIX SCARS FROM PRODUCTION<br><span class="accent">AGENT-MEMORY EVAL, FROM SOMEONE BLEEDING THERE.</span></h1><p class="blog-byline">GUEST · CHUNXIAO WANG, ASSAY / NAUTILUS-COMPASS · EDITED AND PUBLISHED BY XERJ ENGINEERING · 2026-10-07</p>diff --git a/landing/blog/index.html b/landing/blog/index.htmlindex 2da47adcc..816cd12f5 100644--- a/landing/blog/index.html+++ b/landing/blog/index.html@@ -10,7 +10,7 @@<link href="https://fonts.googleapis.com/css2?family=Big+Shoulders+Display:wght@400;700;800;900&family=Inter:wght@400;500;600;700&family=IBM+Plex+Sans:wght@400;500;600;700&family=JetBrains+Mono:wght@400;500;700&display=swap" rel="stylesheet"><link rel="icon" href="data:image/svg+xml,<svg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 100 100'><rect width='100' height='100' fill='%230b0b0d'/><text x='50' y='68' font-size='60' font-family='Big Shoulders Display,sans-serif' font-weight='900' fill='%23ffc400' text-anchor='middle'>X</text></svg>"><link rel="stylesheet" href="/style.css?v=if9">-<link rel="stylesheet" href="/blog.css?v=1">+<link rel="stylesheet" href="/blog.css?v=4"><!-- seo:begin — generated by scripts/seo/fix_heads.py · do not hand-edit --><link rel="canonical" href="https://xerj.org/blog/"><meta property="og:site_name" content="XERJ.ai">@@ -131,35 +131,7 @@<ol class="blog-list"><li><a class="bl-card" href="/blog/the-exploit-hub">- <!-- A/B eval, sealed ground truth: every figure from the run log- (/root/eval-holdout/EVAL-REPORT.md). Bars share one- denominator per row; the adds-~0 case is counted, not- dropped. -->- <div class="bl-preview" aria-hidden="true">- <div class="blp-kicker">CYBER-EXPLOITS CORPUS · A/B EVAL · SEALED TRUTH · 3 FRESH CVES</div>- <div class="blp-row">- <div class="blp-ds">NEEDLE POCS</div>- <div class="blp-pair">- <span class="blp-num">16 <i>needle docs, top 27</i></span><span class="blp-bar a" style="width:59.3%"></span>- <span class="blp-num">0 <i>shipped client, same query</i></span><span class="blp-bar b" style="width:1%"></span>- </div>- </div>- <div class="blp-row">- <div class="blp-ds">COST</div>- <div class="blp-pair">- <span class="blp-num">23 s <i>1.09M docs, cited</i></span><span class="blp-bar a" style="width:100%"></span>- <span class="blp-num">2.1 s <i>clone+grep, repos known</i></span><span class="blp-bar b" style="width:9.1%"></span>- </div>- </div>- <div class="blp-row">- <div class="blp-ds">TALLY</div>- <div class="blp-pair">- <span class="blp-num">2 of 3 attribution <span class="blp-chip">1 of 3 patch</span></span><span class="blp-bar a" style="width:66.7%"></span>- <span class="blp-num">1 adds ~0 <i>counted</i></span><span class="blp-bar b" style="width:33.3%"></span>- </div>- </div>- <p class="blp-foot">group 37.6 GB · exploit-pocs 1,093,749 docs · eval caught client defect #1238</p>- </div>+ <img class="bl-thumb" src="/blog/covers/the-exploit-hub-thumb.webp" width="640" height="336" alt="Cover art: a network of small square nodes; a yellow hub linked to 36 yellow nodes, an ink node with one link, and an empty square frame with none" loading="lazy" decoding="async"><div class="bl-meta"><p class="bl-date"><time datetime="2026-10-08">2026-10-08</time> · CYBER-EXPLOITS CORPUS · EVAL</p><h2 class="bl-title">The exploit hub, tested on three fresh CVEs</h2>@@ -171,35 +143,7 @@</li><li><a class="bl-card" href="/blog/eval-scars-from-agent-memory">- <!-- Guest post: every figure is Chunxiao Wang's, from issues- #1118/#1138; the only XERJ-side number is the §05 registry- count (0 of 99 live + 0 of 132 backlog, validate_hub.py- 2026-10-07). Bars share one denominator per row. -->- <div class="bl-preview" aria-hidden="true">- <div class="blp-kicker">GUEST · AGENT MEMORY · RETRIEVAL EVAL · ISSUES #1118 + #1138</div>- <div class="blp-row">- <div class="blp-ds">EMBEDDER BAKEOFF</div>- <div class="blp-pair">- <span class="blp-num">+0.46pt <i>MTEB pick, r@5</i></span><span class="blp-bar b" style="width:14.4%"></span>- <span class="blp-num">+2.75pt <i>incumbent, r@1 zh</i></span><span class="blp-bar a" style="width:85.7%"></span>- </div>- </div>- <div class="blp-row">- <div class="blp-ds">THE BASELINE</div>- <div class="blp-pair">- <span class="blp-num">0.87 <i>classifier</i></span><span class="blp-bar a" style="width:100%"></span>- <span class="blp-num">0.78 <i>always-yes, n=60</i></span><span class="blp-bar b" style="width:89.7%"></span>- </div>- </div>- <div class="blp-row">- <div class="blp-ds">UNPAIRED VS PAIRED</div>- <div class="blp-pair">- <span class="blp-num">4× <i>overstated</i></span><span class="blp-bar b" style="width:100%"></span>- <span class="blp-num">1× <i>McNemar</i></span><span class="blp-bar a" style="width:25%"></span>- </div>- </div>- <p class="blp-foot">0 of 99 live corpora are agent trajectories · the gap, checked and confirmed</p>- </div>+ <img class="bl-thumb" src="/blog/covers/eval-scars-from-agent-memory-thumb.webp" width="640" height="336" alt="Cover art: six clusters of diagonal scratch marks made of small squares, five in ink and one in yellow" loading="lazy" decoding="async"><div class="bl-meta"><p class="bl-date"><time datetime="2026-10-07">2026-10-07</time> · GUEST · AGENT MEMORY · EVAL</p><h2 class="bl-title">Six eval scars from a production agent-memory system</h2>@@ -219,37 +163,7 @@</li><li><a class="bl-card" href="/blog/the-corpus-hub">- <!-- Bars share one denominator per row (that row's max). Neutral bar =- the plain arm / native baseline, accent bar = the corpus arm.- Figures verbatim from the post and its sources: reference-coding- case study 26,477 vs 9,982 median output tokens / $3.27 vs $1.58;- A/B audit 13 vs 16 dynamically confirmed, 1 FP vs 0, and the row- that decided the verdict — 0 findings traceable to the corpus. -->- <div class="bl-preview" aria-hidden="true">- <div class="blp-kicker">CORPUS HUB · PINNED · LICENCED · SIGNED · OPEN PRS</div>- <div class="blp-row">- <div class="blp-ds">REFCODING OUT TOK</div>- <div class="blp-pair">- <span class="blp-num">26,477 <i>native</i></span><span class="blp-bar b" style="width:100%"></span>- <span class="blp-num">9,982 <i>retrieval</i> <span class="blp-chip">2.7×</span></span><span class="blp-bar a" style="width:37.7%"></span>- </div>- </div>- <div class="blp-row">- <div class="blp-ds">COST / 16 TASKS</div>- <div class="blp-pair">- <span class="blp-num">$3.27 <i>native</i></span><span class="blp-bar b" style="width:100%"></span>- <span class="blp-num">$1.58 <i>retrieval</i> <span class="blp-chip">2.1×</span></span><span class="blp-bar a" style="width:48.3%"></span>- </div>- </div>- <div class="blp-row">- <div class="blp-ds">SECURITY A/B</div>- <div class="blp-pair">- <span class="blp-num">13 + 1 FP <i>plain</i></span><span class="blp-bar b" style="width:81%"></span>- <span class="blp-num">16 + 0 FP <i>corpus</i></span><span class="blp-bar a" style="width:100%"></span>- </div>- </div>- <p class="blp-foot">4 reference corpora + 1 signed pack · 0 corpus-traceable findings · the tie, published</p>- </div>+ <img class="bl-thumb" src="/blog/covers/the-corpus-hub-thumb.webp" width="640" height="336" alt="Cover art: a 10 by 10 wall of square-matrix pack tiles in shades of ink, the first four solid yellow and the fifth framed" loading="lazy" decoding="async"><div class="bl-meta"><p class="bl-date"><time datetime="2026-10-01">2026-10-01</time> · CORPUS HUB · AGENT CONTEXT</p><h2 class="bl-title">The Corpus Hub: pinned, licenced, signed, and contributed by pull request</h2>@@ -267,45 +181,7 @@</li><li><a class="bl-card" href="/blog/the-rc80-gates">- <!-- Bar widths are normalized WITHIN each row against that row's own- max (the card mixes units — F1, a count, an ECE — so one global- denominator would be meaningless). Neutral bar = the bar or the- before, accent bar = the measured value, verbatim from the post:- /_ask 0.9975 vs bar 0.9 · map 400s 20/30 -> 0/30 · ECE 0.3109- raw -> 0.0088 isotonic · judge SciFact 0.7045 -> 0.5967 (the row- we lose, kept in). -->- <div class="bl-preview" aria-hidden="true">- <div class="blp-kicker">RELEASE GATES · TEN FEATURES · EIGHT GATES</div>- <div class="blp-row">- <div class="blp-ds">/_ASK F1</div>- <div class="blp-pair">- <span class="blp-num">0.9975 <i>measured</i></span><span class="blp-bar a" style="width:100%"></span>- <span class="blp-num b">0.9000 <i>the bar</i></span><span class="blp-bar b" style="width:90%"></span>- </div>- </div>- <div class="blp-row">- <div class="blp-ds">MAP 400S</div>- <div class="blp-pair">- <span class="blp-num">20/30 <i>guessed</i></span><span class="blp-bar b" style="width:100%"></span>- <span class="blp-num">0/30 <i>map-informed</i></span><span class="blp-bar a" style="width:2%; min-width:3px"></span>- </div>- </div>- <div class="blp-row">- <div class="blp-ds">FIQA ECE</div>- <div class="blp-pair">- <span class="blp-num">0.3109 <i>raw</i></span><span class="blp-bar b" style="width:100%"></span>- <span class="blp-num">0.0088 <i>isotonic</i></span><span class="blp-bar a" style="width:2.8%"></span>- </div>- </div>- <div class="blp-row">- <div class="blp-ds">JUDGE</div>- <div class="blp-pair">- <span class="blp-num">0.7045 <i>hybrid</i></span><span class="blp-bar a" style="width:100%"></span>- <span class="blp-num b">0.5967 <i>+ judge</i> <span class="blp-chip">−0.108</span></span><span class="blp-bar b" style="width:84.7%"></span>- </div>- </div>- <p class="blp-foot">3 PASS · 3 FAIL · 1 SPLIT · four issues filed (#1091–#1094) · runs 2026-09-30</p>- </div>+ <img class="bl-thumb" src="/blog/covers/the-rc80-gates-thumb.webp" width="640" height="336" alt="Cover art: ten gates built from squares in two rows: three yellow with check marks, three ink with crosses, one half yellow and three faint" loading="lazy" decoding="async"><div class="bl-meta"><p class="bl-date"><time datetime="2026-09-30">2026-09-30</time> · RELEASE GATES · ENGINEERING</p><h2 class="bl-title">The rc.80 gates: three passed outright, three failed, one split</h2>@@ -322,35 +198,7 @@</li><li><a class="bl-card" href="/blog/one-vulnerability-three-names">- <!-- Bar widths share ONE denominator, 4,107 envelopes (the max across- all rows), so lengths compare across rows. Neutral bar = what the- two sources filed, accent bar = what identity resolution made of- it. Figures verbatim from the post: 4,107 envelopes → 1,950- records, 1,221 of them merged from both sources. -->- <div class="bl-preview" aria-hidden="true">- <div class="blp-kicker">CORPUS PACKS · 2 SOURCES · UNION-FIND · ED25519</div>- <div class="blp-row">- <div class="blp-ds">FILED</div>- <div class="blp-pair">- <span class="blp-num">4,107 <i>envelopes</i></span><span class="blp-bar b" style="width:100%"></span>- <span class="blp-num">osv.dev + RustSec</span><span class="blp-bar b" style="width:100%"></span>- </div>- </div>- <div class="blp-row">- <div class="blp-ds">RESOLVED</div>- <div class="blp-pair">- <span class="blp-num">1,950 <i>records</i> <span class="blp-chip">−52.5%</span></span><span class="blp-bar a" style="width:47.5%"></span>- <span class="blp-num">1,221 <i>from both</i></span><span class="blp-bar a" style="width:29.7%"></span>- </div>- </div>- <div class="blp-row">- <div class="blp-ds">SIGNED</div>- <div class="blp-pair">- <span class="blp-num">ed25519 over SHA256SUMS <span class="blp-chip">prior art: 0 of 4</span></span><span class="blp-bar a" style="width:100%"></span>- </div>- </div>- <p class="blp-foot">daily rebuild · signed releases pack-rust-vulns-* · cold build ~34 s</p>- </div>+ <img class="bl-thumb" src="/blog/covers/one-vulnerability-three-names-thumb.webp" width="640" height="336" alt="Cover art: three outlined boxes trailing lines of small squares that converge on one square-matrix disc with a yellow seal at its centre" loading="lazy" decoding="async"><div class="bl-meta"><p class="bl-date"><time datetime="2026-09-27">2026-09-27</time> · CORPUS PACKS · SECURITY DATA</p><h2 class="bl-title">One vulnerability, three names, and a pack we sign</h2>@@ -367,44 +215,7 @@</li><li><a class="bl-card" href="/blog/cutting-the-index-28-percent">- <!-- Bar widths share ONE denominator, 2,449 kB (the max across all rows,- .post before), so lengths compare across rows. Neutral bar = the- stage-0 baseline, accent bar = after that family's format change.- Figures verbatim from the post: .post 2,449/2,365 kB · .seg- 1,671/449 kB · .dv 390/355 kB · .meta 200.8/13.8 kB (the 2 px- sliver is the real ratio). -->- <div class="bl-preview" aria-hidden="true">- <div class="blp-kicker">ON-DISK SIZE · 4 FILE FORMATS · 100K DOCS</div>- <div class="blp-row">- <div class="blp-ds">.POST</div>- <div class="blp-pair">- <span class="blp-num">2,449 kB <i>before</i></span><span class="blp-bar b" style="width:100%"></span>- <span class="blp-num">2,365 kB <i>after</i></span><span class="blp-bar a" style="width:96.6%"></span>- </div>- </div>- <div class="blp-row">- <div class="blp-ds">.SEG</div>- <div class="blp-pair">- <span class="blp-num">1,671 kB <i>before</i></span><span class="blp-bar b" style="width:68.2%"></span>- <span class="blp-num">449 kB <i>after</i> <span class="blp-chip">−73.1%</span></span><span class="blp-bar a" style="width:18.3%"></span>- </div>- </div>- <div class="blp-row">- <div class="blp-ds">.DV</div>- <div class="blp-pair">- <span class="blp-num">390 kB <i>before</i></span><span class="blp-bar b" style="width:15.9%"></span>- <span class="blp-num">355 kB <i>after</i></span><span class="blp-bar a" style="width:14.5%"></span>- </div>- </div>- <div class="blp-row">- <div class="blp-ds">.META</div>- <div class="blp-pair">- <span class="blp-num">200.8 kB <i>before</i></span><span class="blp-bar b" style="width:8.2%"></span>- <span class="blp-num">13.8 kB <i>after</i> <span class="blp-chip">−93.1%</span></span><span class="blp-bar a" style="width:0.6%"></span>- </div>- </div>- <p class="blp-foot">stage 1: ~5.52 MB → 3.99 MB (−28%) · committed control pair 5,398,478 → 3,989,254 B</p>- </div>+ <img class="bl-thumb" src="/blog/covers/cutting-the-index-28-percent-thumb.webp" width="640" height="336" alt="Cover art: a large ink block of squares wrapped in a yellow L-shaped band that marks the 28% cut, beside two pairs of thin before-and-after bars" loading="lazy" decoding="async"><div class="bl-meta"><p class="bl-date"><time datetime="2026-09-27">2026-09-27</time> · STORAGE · ENGINEERING</p><h2 class="bl-title">Cutting the index 28%: dissect first, write Rust second</h2>@@ -419,35 +230,7 @@</li><li><a class="bl-card" href="/blog/jev-vs-a-bm25-vote">- <!-- Bar widths share ONE denominator, 0.9880 (the max across all rows),- so lengths compare across rows. Figures verbatim from the post:- SMS 0.9840 hybrid / 0.9880 Jev · B77 0.9140 hybrid / 0.8240 Jev ·- DBpedia 0.9125 semantic / 0.9800 Jev (the row we lose). -->- <div class="bl-preview" aria-hidden="true">- <div class="blp-kicker">XERJ vs JEV · ACCURACY · 500 DECISIONS EACH</div>- <div class="blp-row">- <div class="blp-ds">SMS SPAM</div>- <div class="blp-pair">- <span class="blp-num">0.9840 <i>hybrid</i></span><span class="blp-bar a" style="width:99.6%"></span>- <span class="blp-num b">0.9880 <i>Jev</i></span><span class="blp-bar b" style="width:100%"></span>- </div>- </div>- <div class="blp-row">- <div class="blp-ds">BANKING77</div>- <div class="blp-pair">- <span class="blp-num">0.9140 <i>hybrid</i> <span class="blp-chip">+0.090</span></span><span class="blp-bar a" style="width:92.5%"></span>- <span class="blp-num b">0.8240 <i>Jev</i></span><span class="blp-bar b" style="width:83.4%"></span>- </div>- </div>- <div class="blp-row">- <div class="blp-ds">DBPEDIA</div>- <div class="blp-pair">- <span class="blp-num">0.9125 <i>semantic</i></span><span class="blp-bar a" style="width:92.4%"></span>- <span class="blp-num b">0.9800 <i>Jev</i></span><span class="blp-bar b" style="width:99.2%"></span>- </div>- </div>- <p class="blp-foot">$0 · 0 tokens (XERJ, local) vs $0.0709 per 1k decisions (Jev, Banking77)</p>- </div>+ <img class="bl-thumb" src="/blog/covers/jev-vs-a-bm25-vote-thumb.webp" width="640" height="336" alt="Cover art: two grids of 500 squares, filled for correct answers; the ink grid has a band of yellow squares where it beats the blue grid" loading="lazy" decoding="async"><div class="bl-meta"><p class="bl-date"><time datetime="2026-09-22">2026-09-22</time> · BENCHMARKS · DECISIONS</p><h2 class="bl-title">A BM25 vote ties Jev on spam and beats it on intent</h2>@@ -463,38 +246,7 @@</li><li><a class="bl-card" href="/blog/does-xerj-beat-jev">- <!-- Preview bar widths are normalized against ONE global denominator,- 0.7410 (the max across all rows), so lengths are comparable ACROSS- rows — the same convention as the og card and the homepage banner.- SciFact 0.6993/0.7410 · NFCorpus 0.3448/0.3312 · FiQA 0.2382/0.3638.- Arm suffixes name which arm each number is, verbatim from the post:- XERJ side hybrid on SciFact/NFCorpus, BM25 on FiQA (hybrid not run);- Jev side is BM25 top-30 + Jev throughout. The og raster stays clean. -->- <div class="bl-preview" aria-hidden="true">- <div class="blp-kicker">XERJ vs JEV · nDCG@10 · 1,271 QUERIES</div>- <div class="blp-row">- <div class="blp-ds">SCIFACT</div>- <div class="blp-pair">- <span class="blp-num">0.6993 <i>hybrid</i></span><span class="blp-bar a" style="width:94.4%"></span>- <span class="blp-num b">0.7410 <i>BM25+Jev</i></span><span class="blp-bar b" style="width:100%"></span>- </div>- </div>- <div class="blp-row">- <div class="blp-ds">NFCORPUS</div>- <div class="blp-pair">- <span class="blp-num">0.3448 <i>hybrid</i></span><span class="blp-bar a" style="width:46.5%"></span>- <span class="blp-num b">0.3312 <i>BM25+Jev</i></span><span class="blp-bar b" style="width:44.7%"></span>- </div>- </div>- <div class="blp-row">- <div class="blp-ds">FIQA</div>- <div class="blp-pair">- <span class="blp-num">0.2382 <i>BM25</i></span><span class="blp-bar a" style="width:32.1%"></span>- <span class="blp-num b">0.3638 <i>BM25+Jev</i> <span class="blp-chip">+0.126</span></span><span class="blp-bar b" style="width:49.1%"></span>- </div>- </div>- <p class="blp-foot">$0.6375 · 15,178,758 tokens (JEV, FiQA) vs $0 · 0 tokens (XERJ local)</p>- </div>+ <img class="bl-thumb" src="/blog/covers/does-xerj-beat-jev-thumb.webp" width="640" height="336" alt="Cover art: three pairs of square-matrix bars, ink for XERJ and blue for BM25+Jev, over one long blue bar and a single yellow square for the zero-cost local run" loading="lazy" decoding="async"><div class="bl-meta"><p class="bl-date"><time datetime="2026-09-21">2026-09-21</time> · BENCHMARKS · JEV</p><h2 class="bl-title">Does XERJ beat JEV? On the bill, outright. On FiQA, no.</h2>diff --git a/landing/blog/jev-vs-a-bm25-vote.html b/landing/blog/jev-vs-a-bm25-vote.htmlindex 5412ad5e8..93a2dc566 100644--- a/landing/blog/jev-vs-a-bm25-vote.html+++ b/landing/blog/jev-vs-a-bm25-vote.html@@ -1,5 +1,5 @@<!DOCTYPE html>-<html lang="en">+<html lang="en" data-theme="day"><head><meta charset="UTF-8"><meta name="viewport" content="width=device-width, initial-scale=1.0">@@ -9,7 +9,7 @@<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin><link href="https://fonts.googleapis.com/css2?family=Big+Shoulders+Display:wght@400;700;800;900&family=Inter:wght@400;500;600;700&family=IBM+Plex+Sans:wght@400;500;600&family=JetBrains+Mono:wght@400;500;700&display=swap" rel="stylesheet"><link rel="stylesheet" href="/style.css?v=if9">-<link rel="stylesheet" href="/blog.css?v=1">+<link rel="stylesheet" href="/blog.css?v=4"><!-- seo:begin — generated by scripts/seo/fix_heads.py · do not hand-edit --><link rel="canonical" href="https://xerj.org/blog/jev-vs-a-bm25-vote"><meta property="og:site_name" content="XERJ.ai">@@ -17,15 +17,15 @@<meta property="og:title" content="A BM25 vote ties Jev on spam and beats it on intent"><meta property="og:description" content="500 decisions each on SMS Spam and Banking77: a tie on spam (McNemar p=0.75), a 9-point retrieval win on 77-way intent at zero cost, and one dataset Jev wins."><meta property="og:url" content="https://xerj.org/blog/jev-vs-a-bm25-vote">-<meta property="og:image" content="https://xerj.org/og/xerj-card.png">+<meta property="og:image" content="https://xerj.org/og/blog-jev-vs-a-bm25-vote.png"><meta property="og:image:width" content="1200"><meta property="og:image:height" content="630">-<meta property="og:image:alt" content="XERJ.ai — the search engine built for AI agents">+<meta property="og:image:alt" content="Cover: “A BM25 vote ties Jev on spam and beats it on intent” in display type beside the post’s square-matrix art, labelled with its figures. The art: two grids of 500 squares, filled for correct answers; the ink grid has a band of yellow squares where it beats the blue grid"><meta property="og:locale" content="en_US"><meta name="twitter:card" content="summary_large_image"><meta name="twitter:title" content="A BM25 vote ties Jev on spam and beats it on intent"><meta name="twitter:description" content="500 decisions each on SMS Spam and Banking77: a tie on spam (McNemar p=0.75), a 9-point retrieval win on 77-way intent at zero cost, and one dataset Jev wins.">-<meta name="twitter:image" content="https://xerj.org/og/xerj-card.png">+<meta name="twitter:image" content="https://xerj.org/og/blog-jev-vs-a-bm25-vote.png"><script type="application/ld+json">{"@context": "https://schema.org",@@ -60,7 +60,7 @@"headline": "A BM25 vote ties Jev on spam and beats it on intent","description": "500 decisions each on SMS Spam and Banking77: a tie on spam (McNemar p=0.75), a 9-point retrieval win on 77-way intent at zero cost, and one dataset Jev wins.","url": "https://xerj.org/blog/jev-vs-a-bm25-vote",- "image": "https://xerj.org/og/xerj-card.png",+ "image": "https://xerj.org/og/blog-jev-vs-a-bm25-vote.png","inLanguage": "en","datePublished": "2026-09-22","dateModified": "2026-09-27",@@ -130,6 +130,7 @@<main><div class="uc-hero"><p class="blog-crumb"><a href="/blog/">← ALL POSTS · /BLOG</a></p>+ <figure class="blog-cover"><img src="/blog/covers/jev-vs-a-bm25-vote-hero.webp" width="1600" height="640" alt="Cover art: two grids of 500 squares, filled for correct answers; the ink grid has a band of yellow squares where it beats the blue grid" fetchpriority="high" decoding="async"></figure><div class="uc-eyebrow">BENCHMARKS · DECISIONS · SMS SPAM + BANKING77 · RUNS 2026-09-21</div><h1>A BM25 VOTE TIES JEV ON SPAM<br><span class="accent">AND BEATS IT ON INTENT</span></h1><p class="blog-byline">XERJ ENGINEERING · PUBLISHED 2026-09-22 · UPDATED 2026-09-27</p>diff --git a/landing/blog/one-vulnerability-three-names.html b/landing/blog/one-vulnerability-three-names.htmlindex 414691825..826adf3f7 100644--- a/landing/blog/one-vulnerability-three-names.html+++ b/landing/blog/one-vulnerability-three-names.html@@ -1,5 +1,5 @@<!DOCTYPE html>-<html lang="en">+<html lang="en" data-theme="day"><head><meta charset="UTF-8"><meta name="viewport" content="width=device-width, initial-scale=1.0">@@ -9,7 +9,7 @@<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin><link href="https://fonts.googleapis.com/css2?family=Big+Shoulders+Display:wght@400;700;800;900&family=Inter:wght@400;500;600;700&family=IBM+Plex+Sans:wght@400;500;600&family=JetBrains+Mono:wght@400;500;700&display=swap" rel="stylesheet"><link rel="stylesheet" href="/style.css?v=if9">-<link rel="stylesheet" href="/blog.css?v=1">+<link rel="stylesheet" href="/blog.css?v=4"><!-- seo:begin — generated by scripts/seo/fix_heads.py · do not hand-edit --><link rel="canonical" href="https://xerj.org/blog/one-vulnerability-three-names"><meta property="og:site_name" content="XERJ.ai">@@ -17,15 +17,15 @@<meta property="og:title" content="One vulnerability, three names — and a pack we sign"><meta property="og:description" content="The same Rust vulnerability arrives as RUSTSEC, GHSA and CVE. We collapsed 4,107 advisories to 1,950 identity-resolved records, packed them, and signed the pack — because no vulnerability database we surveyed does."><meta property="og:url" content="https://xerj.org/blog/one-vulnerability-three-names">-<meta property="og:image" content="https://xerj.org/og/xerj-card.png">+<meta property="og:image" content="https://xerj.org/og/blog-one-vulnerability-three-names.png"><meta property="og:image:width" content="1200"><meta property="og:image:height" content="630">-<meta property="og:image:alt" content="XERJ.ai — the search engine built for AI agents">+<meta property="og:image:alt" content="Cover: “One vulnerability, three names” in display type beside the post’s square-matrix art, labelled with its figures. The art: three outlined boxes trailing lines of small squares that converge on one square-matrix disc with a yellow seal at its centre"><meta property="og:locale" content="en_US"><meta name="twitter:card" content="summary_large_image"><meta name="twitter:title" content="One vulnerability, three names — and a pack we sign"><meta name="twitter:description" content="The same Rust vulnerability arrives as RUSTSEC, GHSA and CVE. We collapsed 4,107 advisories to 1,950 identity-resolved records, packed them, and signed the pack — because no vulnerability database we surveyed does.">-<meta name="twitter:image" content="https://xerj.org/og/xerj-card.png">+<meta name="twitter:image" content="https://xerj.org/og/blog-one-vulnerability-three-names.png"><script type="application/ld+json">{"@context": "https://schema.org",@@ -60,7 +60,7 @@"headline": "One vulnerability, three names — and a pack we sign","description": "The same Rust vulnerability arrives as RUSTSEC, GHSA and CVE. We collapsed 4,107 advisories to 1,950 identity-resolved records, packed them, and signed the pack — because no vulnerability database we surveyed does.","url": "https://xerj.org/blog/one-vulnerability-three-names",- "image": "https://xerj.org/og/xerj-card.png",+ "image": "https://xerj.org/og/blog-one-vulnerability-three-names.png","inLanguage": "en","datePublished": "2026-09-27","dateModified": "2026-09-27",@@ -130,6 +130,7 @@<main><div class="uc-hero"><p class="blog-crumb"><a href="/blog/">← ALL POSTS · /BLOG</a></p>+ <figure class="blog-cover"><img src="/blog/covers/one-vulnerability-three-names-hero.webp" width="1600" height="640" alt="Cover art: three outlined boxes trailing lines of small squares that converge on one square-matrix disc with a yellow seal at its centre" fetchpriority="high" decoding="async"></figure><div class="uc-eyebrow">CORPUS PACKS · IDENTITY RESOLUTION · ED25519 · BUILT 2026-09-26/27</div><h1>ONE VULNERABILITY, THREE NAMES,<br><span class="accent">AND A PACK WE SIGN</span></h1><p class="blog-byline">XERJ ENGINEERING · PUBLISHED 2026-09-27</p>diff --git a/landing/blog/the-corpus-hub.html b/landing/blog/the-corpus-hub.htmlindex 4baa52a8a..2b76acca0 100644--- a/landing/blog/the-corpus-hub.html+++ b/landing/blog/the-corpus-hub.html@@ -1,5 +1,5 @@<!DOCTYPE html>-<html lang="en">+<html lang="en" data-theme="day"><head><meta charset="UTF-8"><meta name="viewport" content="width=device-width, initial-scale=1.0">@@ -9,7 +9,7 @@<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin><link href="https://fonts.googleapis.com/css2?family=Big+Shoulders+Display:wght@400;700;800;900&family=Inter:wght@400;500;600;700&family=IBM+Plex+Sans:wght@400;500;600&family=JetBrains+Mono:wght@400;500;700&display=swap" rel="stylesheet"><link rel="stylesheet" href="/style.css?v=if9">-<link rel="stylesheet" href="/blog.css?v=1">+<link rel="stylesheet" href="/blog.css?v=4"><!-- seo:begin — generated by scripts/seo/fix_heads.py · do not hand-edit --><link rel="canonical" href="https://xerj.org/blog/the-corpus-hub"><meta property="og:site_name" content="XERJ.ai">@@ -17,15 +17,15 @@<meta property="og:title" content="The Corpus Hub: pinned, licenced, signed, contributed by PR"><meta property="og:description" content="A public registry of corpora for AI agents: reference code pinned to SHAs, record packs ed25519-signed, every source licence-reviewed by a human. The measured use case (2.7× fewer output tokens), the tie we published anyway, and how to contribute yours by PR."><meta property="og:url" content="https://xerj.org/blog/the-corpus-hub">-<meta property="og:image" content="https://xerj.org/og/xerj-card.png">+<meta property="og:image" content="https://xerj.org/og/blog-the-corpus-hub.png"><meta property="og:image:width" content="1200"><meta property="og:image:height" content="630">-<meta property="og:image:alt" content="XERJ.ai — the search engine built for AI agents">+<meta property="og:image:alt" content="Cover: “The Corpus Hub” in display type beside the post’s square-matrix art, labelled with its figures. The art: a 10 by 10 wall of square-matrix pack tiles in shades of ink, the first four solid yellow and the fifth framed"><meta property="og:locale" content="en_US"><meta name="twitter:card" content="summary_large_image"><meta name="twitter:title" content="The Corpus Hub: pinned, licenced, signed, contributed by PR"><meta name="twitter:description" content="A public registry of corpora for AI agents: reference code pinned to SHAs, record packs ed25519-signed, every source licence-reviewed by a human. The measured use case (2.7× fewer output tokens), the tie we published anyway, and how to contribute yours by PR.">-<meta name="twitter:image" content="https://xerj.org/og/xerj-card.png">+<meta name="twitter:image" content="https://xerj.org/og/blog-the-corpus-hub.png"><script type="application/ld+json">{"@context": "https://schema.org",@@ -60,7 +60,7 @@"headline": "The Corpus Hub: pinned, licenced, signed, contributed by PR","description": "A public registry of corpora for AI agents: reference code pinned to SHAs, record packs ed25519-signed, every source licence-reviewed by a human. The measured use case (2.7× fewer output tokens), the tie we published anyway, and how to contribute yours by PR.","url": "https://xerj.org/blog/the-corpus-hub",- "image": "https://xerj.org/og/xerj-card.png",+ "image": "https://xerj.org/og/blog-the-corpus-hub.png","inLanguage": "en","datePublished": "2026-10-01","dateModified": "2026-10-06",@@ -130,6 +130,7 @@<main><div class="uc-hero"><p class="blog-crumb"><a href="/blog/">← ALL POSTS · /BLOG</a></p>+ <figure class="blog-cover"><img src="/blog/covers/the-corpus-hub-hero.webp" width="1600" height="640" alt="Cover art: a 10 by 10 wall of square-matrix pack tiles in shades of ink, the first four solid yellow and the fifth framed" fetchpriority="high" decoding="async"></figure><div class="uc-eyebrow">CORPUS HUB · PINNED · LICENCED · SIGNED · BUILT 2026-10-01</div><h1>THE CORPUS HUB: PINNED, LICENCED, SIGNED,<br><span class="accent">AND CONTRIBUTED BY PULL REQUEST</span></h1><p class="blog-byline">XERJ ENGINEERING · PUBLISHED 2026-10-01</p>diff --git a/landing/blog/the-exploit-hub.html b/landing/blog/the-exploit-hub.htmlindex 408b15ad7..b0576e215 100644--- a/landing/blog/the-exploit-hub.html+++ b/landing/blog/the-exploit-hub.html@@ -1,5 +1,5 @@<!DOCTYPE html>-<html lang="en">+<html lang="en" data-theme="day"><head><meta charset="UTF-8"><meta name="viewport" content="width=device-width, initial-scale=1.0">@@ -9,7 +9,7 @@<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin><link href="https://fonts.googleapis.com/css2?family=Big+Shoulders+Display:wght@400;700;800;900&family=Inter:wght@400;500;600;700&family=IBM+Plex+Sans:wght@400;500;600&family=JetBrains+Mono:wght@400;500;700&display=swap" rel="stylesheet"><link rel="stylesheet" href="/style.css?v=if9">-<link rel="stylesheet" href="/blog.css?v=1">+<link rel="stylesheet" href="/blog.css?v=4"><!-- seo:begin — generated by scripts/seo/fix_heads.py · do not hand-edit --><link rel="canonical" href="https://xerj.org/blog/the-exploit-hub"><meta property="og:site_name" content="XERJ.ai">@@ -17,15 +17,15 @@<meta property="og:title" content="The exploit hub, tested on three fresh CVEs"><meta property="og:description" content="We ran the exploit corpus against three CVEs it had never seen: one query named the exact route and field for one, an in-corpus 2026 exploit named the exact denylist segments the vendor shipped for another, and the third proved the corpus adds nothing when its stratum is empty. Losses included."><meta property="og:url" content="https://xerj.org/blog/the-exploit-hub">-<meta property="og:image" content="https://xerj.org/og/xerj-card.png">+<meta property="og:image" content="https://xerj.org/og/blog-the-exploit-hub.png"><meta property="og:image:width" content="1200"><meta property="og:image:height" content="630">-<meta property="og:image:alt" content="XERJ.ai — the search engine built for AI agents">+<meta property="og:image:alt" content="Cover: “The exploit hub, tested on three fresh CVEs” in display type beside the post’s square-matrix art, labelled with its figures. The art: a network of small square nodes; a yellow hub linked to 36 yellow nodes, an ink node with one link, and an empty square frame with none"><meta property="og:locale" content="en_US"><meta name="twitter:card" content="summary_large_image"><meta name="twitter:title" content="The exploit hub, tested on three fresh CVEs"><meta name="twitter:description" content="We ran the exploit corpus against three CVEs it had never seen: one query named the exact route and field for one, an in-corpus 2026 exploit named the exact denylist segments the vendor shipped for another, and the third proved the corpus adds nothing when its stratum is empty. Losses included.">-<meta name="twitter:image" content="https://xerj.org/og/xerj-card.png">+<meta name="twitter:image" content="https://xerj.org/og/blog-the-exploit-hub.png"><script type="application/ld+json">{"@context": "https://schema.org",@@ -60,7 +60,7 @@"headline": "The exploit hub, tested on three fresh CVEs","description": "We ran the exploit corpus against three CVEs it had never seen: one query named the exact route and field for one, an in-corpus 2026 exploit named the exact denylist segments the vendor shipped for another, and the third proved the corpus adds nothing when its stratum is empty. Losses included.","url": "https://xerj.org/blog/the-exploit-hub",- "image": "https://xerj.org/og/xerj-card.png",+ "image": "https://xerj.org/og/blog-the-exploit-hub.png","inLanguage": "en","datePublished": "2026-10-08","dateModified": "2026-10-08",@@ -130,6 +130,7 @@<main><div class="uc-hero"><p class="blog-crumb"><a href="/blog/">← ALL POSTS · /BLOG</a></p>+ <figure class="blog-cover"><img src="/blog/covers/the-exploit-hub-hero.webp" width="1600" height="640" alt="Cover art: a network of small square nodes; a yellow hub linked to 36 yellow nodes, an ink node with one link, and an empty square frame with none" fetchpriority="high" decoding="async"></figure><div class="uc-eyebrow">CYBER-EXPLOITS CORPUS · A/B EVAL · SEALED GROUND TRUTH · RUN 2026-10-08</div><h1>THE EXPLOIT HUB, TESTED ON<br><span class="accent">THREE FRESH CVES</span></h1><p class="blog-byline">XERJ ENGINEERING · PUBLISHED 2026-10-08</p>diff --git a/landing/blog/the-rc80-gates.html b/landing/blog/the-rc80-gates.htmlindex a9ae57b45..cdb4b3c46 100644--- a/landing/blog/the-rc80-gates.html+++ b/landing/blog/the-rc80-gates.html@@ -10,7 +10,7 @@<link href="https://fonts.googleapis.com/css2?family=Big+Shoulders+Display:wght@400;700;800;900&family=Inter:wght@400;500;600;700&family=IBM+Plex+Sans:wght@400;500;600;700&family=JetBrains+Mono:wght@400;500;700&display=swap" rel="stylesheet"><link rel="icon" href="data:image/svg+xml,<svg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 100 100'><rect width='100' height='100' fill='%230b0b0d'/><text x='50' y='68' font-size='60' font-family='Big Shoulders Display,sans-serif' font-weight='900' fill='%23ffc400' text-anchor='middle'>X</text></svg>"><link rel="stylesheet" href="/style.css?v=if9">-<link rel="stylesheet" href="/blog.css?v=1">+<link rel="stylesheet" href="/blog.css?v=4"><!-- seo:begin — generated by scripts/seo/fix_heads.py · do not hand-edit --><link rel="canonical" href="https://xerj.org/blog/the-rc80-gates"><meta property="og:site_name" content="XERJ.ai">@@ -18,15 +18,15 @@<meta property="og:title" content="The rc.80 gates: three passed outright, three failed, one split"><meta property="og:description" content="Ten features, eight measured gates: /_ask macro-F1 0.9975, xerj_map 20/30 to 0/30, calibration ECE 0.0088 pass; the judge, the flywheel and the decide head fail their bars, numbers published."><meta property="og:url" content="https://xerj.org/blog/the-rc80-gates">-<meta property="og:image" content="https://xerj.org/og/xerj-card.png">+<meta property="og:image" content="https://xerj.org/og/blog-the-rc80-gates.png"><meta property="og:image:width" content="1200"><meta property="og:image:height" content="630">-<meta property="og:image:alt" content="XERJ.ai — the search engine built for AI agents">+<meta property="og:image:alt" content="Cover: “The rc.80 gates” in display type beside the post’s square-matrix art, labelled with its figures. The art: ten gates built from squares in two rows: three yellow with check marks, three ink with crosses, one half yellow and three faint"><meta property="og:locale" content="en_US"><meta name="twitter:card" content="summary_large_image"><meta name="twitter:title" content="The rc.80 gates: three passed outright, three failed, one split"><meta name="twitter:description" content="Ten features, eight measured gates: /_ask macro-F1 0.9975, xerj_map 20/30 to 0/30, calibration ECE 0.0088 pass; the judge, the flywheel and the decide head fail their bars, numbers published.">-<meta name="twitter:image" content="https://xerj.org/og/xerj-card.png">+<meta name="twitter:image" content="https://xerj.org/og/blog-the-rc80-gates.png"><script type="application/ld+json">{"@context": "https://schema.org",@@ -61,7 +61,7 @@"headline": "The rc.80 gates: three passed outright, three failed, one split","description": "Ten features, eight measured gates: /_ask macro-F1 0.9975, xerj_map 20/30 to 0/30, calibration ECE 0.0088 pass; the judge, the flywheel and the decide head fail their bars, numbers published.","url": "https://xerj.org/blog/the-rc80-gates",- "image": "https://xerj.org/og/xerj-card.png",+ "image": "https://xerj.org/og/blog-the-rc80-gates.png","inLanguage": "en","datePublished": "2026-09-30","dateModified": "2026-09-30",@@ -132,6 +132,7 @@<main><div class="uc-hero"><p class="blog-crumb"><a href="/blog/">← ALL POSTS · /BLOG</a></p>+ <figure class="blog-cover"><img src="/blog/covers/the-rc80-gates-hero.webp" width="1600" height="640" alt="Cover art: ten gates built from squares in two rows: three yellow with check marks, three ink with crosses, one half yellow and three faint" fetchpriority="high" decoding="async"></figure><div class="uc-eyebrow">RELEASE GATES · TEN FEATURES · EIGHT GATES · RUNS 2026-09-30</div><h1>THE RC.80 GATES.<br><span class="accent">MEASURED IN PUBLIC, LOSSES LEFT IN.</span></h1><p class="blog-byline">AUTHOR · XERJ ENGINEERING TEAM · PUBLISHED · 2026-09-30</p>diff --git a/landing/og/blog-cutting-the-index-28-percent.png b/landing/og/blog-cutting-the-index-28-percent.pngnew file mode 100644index 000000000..0e47075e2Binary files /dev/null and b/landing/og/blog-cutting-the-index-28-percent.png differdiff --git a/landing/og/blog-does-xerj-beat-jev.png b/landing/og/blog-does-xerj-beat-jev.pngnew file mode 100644index 000000000..7dced2302Binary files /dev/null and b/landing/og/blog-does-xerj-beat-jev.png differdiff --git a/landing/og/blog-eval-scars-from-agent-memory.png b/landing/og/blog-eval-scars-from-agent-memory.pngnew file mode 100644index 000000000..9244ff686Binary files /dev/null and b/landing/og/blog-eval-scars-from-agent-memory.png differdiff --git a/landing/og/blog-jev-vs-a-bm25-vote.png b/landing/og/blog-jev-vs-a-bm25-vote.pngnew file mode 100644index 000000000..23d7650f3Binary files /dev/null and b/landing/og/blog-jev-vs-a-bm25-vote.png differdiff --git a/landing/og/blog-one-vulnerability-three-names.png b/landing/og/blog-one-vulnerability-three-names.pngnew file mode 100644index 000000000..f90ffb2a5Binary files /dev/null and b/landing/og/blog-one-vulnerability-three-names.png differdiff --git a/landing/og/blog-the-corpus-hub.png b/landing/og/blog-the-corpus-hub.pngnew file mode 100644index 000000000..8af7ab946Binary files /dev/null and b/landing/og/blog-the-corpus-hub.png differdiff --git a/landing/og/blog-the-exploit-hub.png b/landing/og/blog-the-exploit-hub.pngnew file mode 100644index 000000000..44d6e8439Binary files /dev/null and b/landing/og/blog-the-exploit-hub.png differdiff --git a/landing/og/blog-the-rc80-gates.png b/landing/og/blog-the-rc80-gates.pngnew file mode 100644index 000000000..716081e1dBinary files /dev/null and b/landing/og/blog-the-rc80-gates.png differdiff --git a/landing/sitemap.xml b/landing/sitemap.xmlindex d8d79c9f5..da4e2bb25 100644--- a/landing/sitemap.xml+++ b/landing/sitemap.xml@@ -334,39 +334,39 @@</url><url><loc>https://xerj.org/blog/</loc>- <lastmod>2026-10-18</lastmod>+ <lastmod>2026-10-20</lastmod></url><url><loc>https://xerj.org/blog/cutting-the-index-28-percent</loc>- <lastmod>2026-10-18</lastmod>+ <lastmod>2026-10-20</lastmod></url><url><loc>https://xerj.org/blog/does-xerj-beat-jev</loc>- <lastmod>2026-10-18</lastmod>+ <lastmod>2026-10-20</lastmod></url><url><loc>https://xerj.org/blog/eval-scars-from-agent-memory</loc>- <lastmod>2026-10-13</lastmod>+ <lastmod>2026-10-20</lastmod></url><url><loc>https://xerj.org/blog/jev-vs-a-bm25-vote</loc>- <lastmod>2026-10-18</lastmod>+ <lastmod>2026-10-20</lastmod></url><url><loc>https://xerj.org/blog/one-vulnerability-three-names</loc>- <lastmod>2026-10-18</lastmod>+ <lastmod>2026-10-20</lastmod></url><url><loc>https://xerj.org/blog/the-corpus-hub</loc>- <lastmod>2026-10-13</lastmod>+ <lastmod>2026-10-20</lastmod></url><url><loc>https://xerj.org/blog/the-exploit-hub</loc>- <lastmod>2026-10-13</lastmod>+ <lastmod>2026-10-20</lastmod></url><url><loc>https://xerj.org/blog/the-rc80-gates</loc>- <lastmod>2026-10-18</lastmod>+ <lastmod>2026-10-20</lastmod></url><url><loc>https://xerj.org/brand</loc>diff --git a/scripts/seo/blog_covers/.gitignore b/scripts/seo/blog_covers/.gitignorenew file mode 100644index 000000000..6f3cd8ffe--- /dev/null+++ b/scripts/seo/blog_covers/.gitignore@@ -0,0 +1 @@+.render-*.htmldiff --git a/scripts/seo/blog_covers/README.md b/scripts/seo/blog_covers/README.mdnew file mode 100644index 000000000..af616bb07--- /dev/null+++ b/scripts/seo/blog_covers/README.md@@ -0,0 +1,89 @@+# `scripts/seo/blog_covers/`: the blog cover series++Every post on xerj.org/blog gets a cover from one series, the **engineering+field-log poster**, printed in the brand book's palette: light paper+(`#f6f4ee`, the day theme the `/blog/` index is served in) on a 1px square+grid, a square-matrix motif drawn from that post's own numbers in ink+(`#0B0B0D`), a mono kicker, the title in Big Shoulders Display 900, an ochre+(`#A06800`) subtitle and the post URL. Brand yellow (`#FFC400`) is earned: a+fill on the one result the post is about, never text. Flat: 1px lines, no+gradients, no grain, no shadows. The og PNG carries a 1px `#d6d4cb` frame so+it keeps its edge on white timelines (X, Slack, LinkedIn). All of it comes from code: HTML + SVG+rendered by headless Chromium. No image model is involved, and the same+inputs give the same pixels.++The og card carries the title and the motif's figure labels, because a+social feed shows it on its own. The in-post hero and the index thumbnail are+**art only**: no title, no subtitle, no labels (`motifs.LABELS = False`), and+the motif's real bounding box is fitted, centred, into the frame. The page+around them already carries the title, so the cover never repeats the header.++| output | size | used for |+|---|---|---|+| `landing/og/blog-<slug>.png` | 1200×630, 96-colour PNG | `og:image`, `twitter:image`, Article JSON-LD `image` |+| `landing/blog/covers/<slug>-hero.webp` | 1600×640, lossless WebP | the cover at the top of the post |+| `landing/blog/covers/<slug>-thumb.webp` | 640×336, WebP q80 | the card thumbnail on `/blog/` (lazy-loaded) |++## Run it++```sh+python3 -m venv /tmp/covers && /tmp/covers/bin/pip install playwright pillow+/tmp/covers/bin/playwright install chromium # or: export CHROME=/usr/bin/google-chrome+/tmp/covers/bin/python scripts/seo/blog_covers/covers.py # every post+/tmp/covers/bin/python scripts/seo/blog_covers/covers.py --only <slug> # one post+/tmp/covers/bin/python scripts/seo/blog_covers/covers.py --contact /tmp/sheet.png+```++This is an offline authoring tool. CI does not run it (the `scripts/seo/`+gates stay stdlib-only). The outputs are committed, and `seo_lint.py` checks+that every `og:image` exists and is 1200×630.++The fonts are vendored in `fonts/` (Big Shoulders Display and JetBrains Mono,+both SIL OFL 1.1, licence texts beside them), so a render does not depend on+what is installed on the machine. They are the families `landing/style.css`+names as `--font-display` and `--font-mono`. The colours are+`landing/style.css`'s day tokens (paper, mute, line, cmp) plus the brand-book+ink, yellow and ochre (`landing/brand.html`), copied into `TOKENS` in `covers.py`.++## Adding a post++1. Add an entry to `POSTS` in `covers.py`: `slug`, `date` (the post's+ published date), `title`, `subtitle` (a phrase lifted from the post),+ `motif`, `alt` (the hero/thumbnail alt: describe the art only, since+ those covers carry no text) and `og_alt` (the og card: its title plus the art).+2. Pick a motif in `motifs.py` or write one: `motif(w, h, rng) -> svg`.+ Every number it draws must appear in the post. Name the source in the+ motif's docstring. `rng` is seeded with the slug and may only decide+ placement (jitter, scatter), never a value.+3. Render, then wire it:+ - `scripts/seo/pagedata.py`: `og_image="og/blog-<slug>.png"` and+ `og_image_alt=` the same text as `og_alt`, then+ `python3 scripts/seo/fix_heads.py --write`;+ - the post: `<figure class="blog-cover"><img src="/blog/covers/<slug>-hero.webp"+ width="1600" height="640" alt="…" fetchpriority="high" decoding="async"></figure>`+ right after the `blog-crumb` line;+ - `landing/blog/index.html`: `<img class="bl-thumb" src="/blog/covers/<slug>-thumb.webp"+ width="640" height="336" alt="…" loading="lazy" decoding="async">` as the+ first child of the post's `bl-card`.+4. Look at the contact sheet before you commit. If a label collides with+ the motif, fix the motif. Do not shrink the type.++## The motifs++| motif | post | what it draws |+|---|---|---|+| `jevbars` | does-xerj-beat-jev | nDCG@10 square-bars per dataset, one denominator; the FiQA bill |+| `vote` | jev-vs-a-bm25-vote | 500 Banking77 decisions per arm, filled = correct, the vote's extra wins in yellow |+| `index28` | cutting-the-index-28-percent | 5.52 MB and 3.99 MB blocks to scale, the 28% cut in yellow; .seg/.meta bars |+| `names` | one-vulnerability-three-names | three advisory ids converging on one signed record |+| `gates` | the-rc80-gates | ten features as gates: pass / fail / split / unmeasured |+| `tiles` | the-corpus-hub | a wall of 100 pack tiles, the launch four in yellow |+| `scars` | eval-scars-from-agent-memory | six scratch marks, one per pattern |+| `network` | the-exploit-hub | PoC documents linked to the three CVEs of the eval |+| `route` | corpus-hub-v2 | 251 packs, the 14 catalogue corpora, a question routed to lz4 |++Inks: yellow fill (with ochre for its labels and hairlines) for the one+result the post is about, grey-blue (day `--z-cmp`) for the comparison series+and never for XERJ, ink for structure, and mute/line for annotation and empty+cells. Every motif is built from squares on a grid: `dot()` draws a filled+cell, `ring()` an empty one.diff --git a/scripts/seo/blog_covers/covers.py b/scripts/seo/blog_covers/covers.pynew file mode 100644index 000000000..4c0e19e81--- /dev/null+++ b/scripts/seo/blog_covers/covers.py@@ -0,0 +1,391 @@+#!/usr/bin/env python3+"""Render the xerj.org blog covers: one "engineering field-log poster" per post.++Each post gets three files, all drawn from code (HTML + SVG, rasterised by+headless Chromium through Playwright), never from an image model:++ landing/og/blog-<slug>.png 1200x630 og:image / twitter:image+ landing/blog/covers/<slug>-hero.webp 1600x640 the cover at the top of the post+ landing/blog/covers/<slug>-thumb.webp 640x336 the card thumbnail on /blog/++Layout, type and colour are fixed for the series; only the square-matrix motif+(scripts/seo/blog_covers/motifs.py) and the words change per post. The words+live in POSTS below. Every number a cover shows must appear in its post.++ python3 scripts/seo/blog_covers/covers.py # all posts+ python3 scripts/seo/blog_covers/covers.py --only the-corpus-hub+ python3 scripts/seo/blog_covers/covers.py --contact /tmp/contact.png++Needs: pip install playwright pillow, plus a Chromium (`playwright install+chromium`, or point CHROME at an installed Chrome/Chromium binary). This is an+offline authoring tool, like the og-blog render note in ../assets/: CI does+not run it, the PNG/WebP outputs are committed. See README.md.+"""++from __future__ import annotations++import argparse+import html+import io+import os+import pathlib+import random+import shutil+import sys++HERE = pathlib.Path(__file__).resolve().parent+REPO = HERE.parent.parent.parent+LANDING = REPO / "landing"+sys.path.insert(0, str(HERE))+import motifs # noqa: E402+from motifs import MOTIFS # noqa: E402++# ── brand tokens: the brand book's print palette on the site's day paper ────+# Paper and rules are landing/style.css html[data-theme='day'] (the theme the+# /blog/ index is served in); ink, yellow and ochre are landing/brand.html.+# Yellow is a FILL only (a square, a bar, the one result the post is about):+# it never carries text on paper. Text and hairlines that need the accent use+# ochre, which holds contrast on paper.+TOKENS = {+ "bg": "#f6f4ee", # day --z-bg (paper)+ "ink": "#0b0b0d", # brand ink+ "mute": "#696762", # day --z-mute+ "faint": "#ebe8df", # grid paper rule+ "line": "#cfcbbf", # day --z-line+ "accent": "#ffc400", # brand yellow, fills only+ "ochre": "#a06800", # brand ochre: accent text and accent strokes+ "cmp": "#6f8aa8", # day --z-cmp, the comparison series, never XERJ+ "frame": "#d6d4cb", # 1px og frame so the card holds its edge on white+}++# ── the series: one entry per post ───────────────────────────────────────────+# title display face, set uppercase; keep it to the post's own title+# subtitle yellow line under the title; a phrase lifted from the post+# date the post's published date (pagedata.py `published`)+# motif key into motifs.MOTIFS; the motif's docstring names its sources+# alt alt text for the hero/thumb and og:image:alt: what the card shows+POSTS: list[dict[str, str]] = [+ dict(slug="does-xerj-beat-jev", date="2026-09-21", motif="jevbars",+ title="Does XERJ beat JEV?",+ subtitle="On the bill, outright. On FiQA, no.",+ alt="Cover art: three pairs of square-matrix bars, ink for XERJ and blue for BM25+Jev, over one long blue bar and a single yellow square for the zero-cost local run",+ og_alt="Cover: “Does XERJ beat JEV?” in display type beside the post’s square-matrix art, labelled with its figures. The art: three pairs of square-matrix bars, ink for XERJ and blue for BM25+Jev, over one long blue bar and a single yellow square for the zero-cost local run"),+ dict(slug="jev-vs-a-bm25-vote", date="2026-09-22", motif="vote",+ title="A BM25 vote ties Jev on spam and beats it on intent",+ subtitle="Count your labels before you pay for a model.",+ alt="Cover art: two grids of 500 squares, filled for correct answers; the ink grid has a band of yellow squares where it beats the blue grid",+ og_alt="Cover: “A BM25 vote ties Jev on spam and beats it on intent” in display type beside the post’s square-matrix art, labelled with its figures. The art: two grids of 500 squares, filled for correct answers; the ink grid has a band of yellow squares where it beats the blue grid"),+ dict(slug="cutting-the-index-28-percent", date="2026-09-27", motif="index28",+ title="Cutting the index 28%",+ subtitle="Dissect first, write Rust second.",+ alt="Cover art: a large ink block of squares wrapped in a yellow L-shaped band that marks the 28% cut, beside two pairs of thin before-and-after bars",+ og_alt="Cover: “Cutting the index 28%” in display type beside the post’s square-matrix art, labelled with its figures. The art: a large ink block of squares wrapped in a yellow L-shaped band that marks the 28% cut, beside two pairs of thin before-and-after bars"),+ dict(slug="one-vulnerability-three-names", date="2026-09-27", motif="names",+ title="One vulnerability, three names",+ subtitle="Checksums prove integrity. They do not prove origin.",+ alt="Cover art: three outlined boxes trailing lines of small squares that converge on one square-matrix disc with a yellow seal at its centre",+ og_alt="Cover: “One vulnerability, three names” in display type beside the post’s square-matrix art, labelled with its figures. The art: three outlined boxes trailing lines of small squares that converge on one square-matrix disc with a yellow seal at its centre"),+ dict(slug="the-rc80-gates", date="2026-09-30", motif="gates",+ title="The rc.80 gates",+ subtitle="Measured in public, losses left in.",+ alt="Cover art: ten gates built from squares in two rows: three yellow with check marks, three ink with crosses, one half yellow and three faint",+ og_alt="Cover: “The rc.80 gates” in display type beside the post’s square-matrix art, labelled with its figures. The art: ten gates built from squares in two rows: three yellow with check marks, three ink with crosses, one half yellow and three faint"),+ dict(slug="the-corpus-hub", date="2026-10-01", motif="tiles",+ title="The Corpus Hub",+ subtitle="Pinned, licenced, signed, contributed by pull request.",+ alt="Cover art: a 10 by 10 wall of square-matrix pack tiles in shades of ink, the first four solid yellow and the fifth framed",+ og_alt="Cover: “The Corpus Hub” in display type beside the post’s square-matrix art, labelled with its figures. The art: a 10 by 10 wall of square-matrix pack tiles in shades of ink, the first four solid yellow and the fifth framed"),+ dict(slug="eval-scars-from-agent-memory", date="2026-10-07", motif="scars",+ title="Six eval scars from production agent memory",+ subtitle="If every query scores zero hits, suspect your mask.",+ alt="Cover art: six clusters of diagonal scratch marks made of small squares, five in ink and one in yellow",+ og_alt="Cover: “Six eval scars from production agent memory” in display type beside the post’s square-matrix art, labelled with its figures. The art: six clusters of diagonal scratch marks made of small squares, five in ink and one in yellow"),+ dict(slug="the-exploit-hub", date="2026-10-08", motif="network",+ title="The exploit hub, tested on three fresh CVEs",+ subtitle="One query, exact attribution, wrong patch.",+ alt="Cover art: a network of small square nodes; a yellow hub linked to 36 yellow nodes, an ink node with one link, and an empty square frame with none",+ og_alt="Cover: “The exploit hub, tested on three fresh CVEs” in display type beside the post’s square-matrix art, labelled with its figures. The art: a network of small square nodes; a yellow hub linked to 36 yellow nodes, an ink node with one link, and an empty square frame with none"),+ dict(slug="corpus-hub-v2", date="2026-10-19", motif="route",+ title="Corpus Hub v2",+ subtitle="Ask the question, get the one pack that answers it.",+ alt="Cover art: a faint wall of 251 small tiles above 14 larger square-matrix tiles; a yellow route runs from the top into one yellow tile and down to a framed node",+ og_alt="Cover: “Corpus Hub v2” in display type beside the post’s square-matrix art, labelled with its figures. The art: a faint wall of 251 small tiles above 14 larger square-matrix tiles; a yellow route runs from the top into one yellow tile and down to a framed node"),+]++# og: title + motif with its labels, for social feeds that show the card alone.+# hero / thumb: art only (no title, no labels); the page around them already+# carries the title. "thumb" renders at og size and is downscaled to THUMB.+SIZES = {"og": (1200, 630), "hero": (1600, 640), "thumb": (1200, 630)}+ART_ONLY = {"hero", "thumb"}+THUMB = (640, 336)+++def font_url(name: str) -> str:+ return (HERE / "fonts" / name).as_uri()+++CSS = """+@font-face { font-family: 'Big Shoulders Display'; font-weight: 100 900;+ src: url('%(display)s') format('truetype'); }+@font-face { font-family: 'JetBrains Mono'; font-weight: 100 800;+ src: url('%(mono)s') format('truetype'); }+* { box-sizing: border-box; margin: 0; padding: 0; }+html, body { background: %(bg)s; }+.card { position: relative; overflow: hidden; background: %(bg)s; color: %(ink)s;+ width: %(W)spx; height: %(H)spx; }+.bgsvg { position: absolute; inset: 0; }+.kicker { position: absolute; left: %(P)spx; top: %(P)spx; right: %(P)spx;+ font: 600 %(kfs)spx/1 'JetBrains Mono', monospace; letter-spacing: .14em;+ color: %(ink)s; display: flex; gap: 14px; align-items: center; }+.kicker .sq { width: 14px; height: 14px; background: %(accent)s; flex: none; }+.kicker .no { margin-left: auto; color: %(mute)s; }+.textcol { position: absolute; left: %(P)spx; top: %(ttop)spx; width: %(tw)spx;+ height: %(th)spx; display: flex; flex-direction: column; justify-content: center; }+h1 { font-family: 'Big Shoulders Display', sans-serif; font-weight: 900;+ text-transform: uppercase; line-height: .9; letter-spacing: .005em;+ color: %(ink)s; font-size: %(tfs)spx; }+.sub { margin-top: 22px; font-family: 'Big Shoulders Display', sans-serif; font-weight: 700;+ text-transform: uppercase; letter-spacing: .02em; line-height: 1.02;+ color: %(ochre)s; font-size: %(sfs)spx; max-width: %(tw)spx; }+.rule { width: 72px; height: 1px; background: %(ink)s; margin: 0 0 22px; }+.motif { position: absolute; left: %(mx)spx; top: %(my)spx; }+.foot { position: absolute; left: %(P)spx; right: %(P)spx; bottom: %(fb)spx;+ border-top: 1px solid %(line)s; padding-top: 14px; display: flex;+ font: 500 %(ffs)spx/1 'JetBrains Mono', monospace; letter-spacing: .08em; color: %(mute)s; }+.foot b { color: %(ink)s; font-weight: 600; }+.foot .r { margin-left: auto; }+svg text { font-family: 'JetBrains Mono', monospace; }+.y { fill: %(accent)s; } .w { fill: %(ink)s; } .b { fill: %(cmp)s; }+.m { fill: %(mute)s; } .f { fill: %(line)s; }+.y-s { stroke: %(ochre)s; } .w-s { stroke: %(ink)s; } .b-s { stroke: %(cmp)s; }+.m-s { stroke: %(mute)s; } .f-s { stroke: %(line)s; }+.bgfill { fill: %(bg)s; } .fa { fill: %(faint)s; }+svg text { paint-order: stroke; stroke: %(bg)s; stroke-width: 5px; stroke-linejoin: round; }+.yo { fill: %(accent)s; stroke: %(ochre)s; stroke-width: 1px; }+.lab { font-size: %(lfs)spx; font-weight: 600; letter-spacing: .1em; fill: %(mute)s; }+.lab.y { fill: %(ochre)s; } .lab.w { fill: %(ink)s; } .lab.b { fill: %(cmp)s; }+.num { font-size: %(nfs)spx; font-weight: 600; letter-spacing: .02em; }+.num.y { fill: %(ochre)s; } .num.b { fill: %(cmp)s; } .num.w { fill: %(ink)s; } .num.m { fill: %(mute)s; }+.tag { font-size: %(gfs)spx; font-weight: 700; letter-spacing: .06em; }+.tag.y { fill: %(ochre)s; } .tag.w { fill: %(ink)s; } .tag.m { fill: %(mute)s; }+.big { font-family: 'Big Shoulders Display', sans-serif; font-weight: 900; font-size: %(bfs)spx; }+"""++FIT_ART_JS = """+() => {+ const svg = document.getElementById('art'), b = document.getElementById('artg').getBBox();+ const pad = 6;+ svg.setAttribute('viewBox', `${b.x - pad} ${b.y - pad} ${b.width + 2 * pad} ${b.height + 2 * pad}`);+ return [b.width, b.height];+}+"""++FIT_JS = """+() => {+ const h1 = document.querySelector('h1'), col = document.querySelector('.textcol');+ let fs = parseFloat(getComputedStyle(h1).fontSize);+ while (col.scrollHeight > col.clientHeight + 1 && fs > 30) { fs -= 2; h1.style.fontSize = fs + 'px'; }+ return fs;+}+"""+++def layout(kind: str) -> dict:+ W, H = SIZES[kind]+ if kind == "hero":+ return dict(W=W, H=H, P=64, kfs=16, ttop=0, tw=0, th=0, tfs=0, sfs=0,+ mx=90, my=70, mw=1140, mh=470, aw=1420, ah=500, fb=40, ffs=15, lfs=13, nfs=14, gfs=12, bfs=64)+ if kind == "thumb":+ return dict(W=W, H=H, P=56, kfs=15, ttop=0, tw=0, th=0, tfs=0, sfs=0,+ mx=90, my=70, mw=980, mh=470, aw=1020, ah=490, fb=40, ffs=15, lfs=12, nfs=13, gfs=11, bfs=54)+ if kind == "og":+ return dict(W=W, H=H, P=56, kfs=15, ttop=104, tw=500, th=410, tfs=92, sfs=30,+ mx=610, my=112, mw=536, mh=400, fb=40, ffs=15, lfs=12, nfs=13, gfs=11, bfs=54)+ return dict(W=W, H=H, P=64, kfs=16, ttop=110, tw=640, th=410, tfs=104, sfs=32,+ mx=790, my=112, mw=746, mh=410, fb=40, ffs=15, lfs=13, nfs=14, gfs=12, bfs=64)+++def background(W: int, H: int, seed: int) -> str:+ """Grid paper (1px squares, the matrix every motif snaps to) and crop marks.+ Flat: no gradient, no grain, no shadow."""+ marks = []+ for (x, y) in [(22, 22), (W - 22, 22), (22, H - 22), (W - 22, H - 22)]:+ marks.append(f'<line x1="{x-9}" y1="{y}" x2="{x+9}" y2="{y}" stroke="{TOKENS["line"]}"/>'+ f'<line x1="{x}" y1="{y-9}" x2="{x}" y2="{y+9}" stroke="{TOKENS["line"]}"/>')+ return f"""<svg class="bgsvg" width="{W}" height="{H}" viewBox="0 0 {W} {H}" xmlns="http://www.w3.org/2000/svg">+<defs>+ <pattern id="grid" width="24" height="24" patternUnits="userSpaceOnUse">+ <path d="M24 0H0V24" fill="none" stroke="{TOKENS["faint"]}" stroke-width="1"/></pattern>+</defs>+<rect width="{W}" height="{H}" fill="url(#grid)"/>+{''.join(marks)}+</svg>"""+++def page(post: dict, kind: str, number: int, total: int) -> str:+ L = layout(kind)+ seed = sum(post["slug"].encode()) % 997+ rng = random.Random(post["slug"] + ":" + kind)+ motifs.LABELS = kind not in ART_ONLY+ try:+ motif = MOTIFS[post["motif"]](L["mw"], L["mh"], rng)+ finally:+ motifs.LABELS = True+ motif_svg = (f'<svg class="motif" width="{L["mw"]}" height="{L["mh"] + 4}" '+ f'viewBox="0 0 {L["mw"]} {L["mh"] + 4}" overflow="visible" xmlns="http://www.w3.org/2000/svg">'+ f'{motif}</svg>')+ css = CSS % {**TOKENS, **L, "display": font_url("BigShouldersDisplay-wght.ttf"),+ "mono": font_url("JetBrainsMono-wght.ttf")}+ e = html.escape+ if kind in ART_ONLY:+ # the motif is drawn in its own (mw x mh) box, then FIT_ART_JS fits its+ # real bounding box, centred, into the (aw x ah) art area+ art = (f'<svg id="art" class="motif" width="{L["aw"]}" height="{L["ah"]}" '+ f'viewBox="0 0 {L["mw"]} {L["mh"]}" preserveAspectRatio="xMidYMid meet" '+ f'xmlns="http://www.w3.org/2000/svg"><g id="artg">{motif}</g></svg>')+ return f"""<!DOCTYPE html><html><head><meta charset="utf-8"><style>{css}</style></head><body>+<div class="card">+{background(L["W"], L["H"], seed)}+{art}+</div></body></html>"""+ return f"""<!DOCTYPE html><html><head><meta charset="utf-8"><style>{css}</style></head><body>+<div class="card">+{background(L["W"], L["H"], seed)}+<div class="kicker"><span class="sq"></span><span>XERJ · ENGINEERING LOG · {e(post["date"])}</span>+<span class="no">No. {number:02d} / {total:02d}</span></div>+<div class="textcol"><div class="rule"></div><h1>{e(post["title"])}</h1><p class="sub">{e(post["subtitle"])}</p></div>+{motif_svg}+<div class="foot"><span><b>xerj.org/blog/{e(post["slug"])}</b></span><span class="r">DRAWN FROM THE POST'S OWN NUMBERS</span></div>+</div></body></html>"""+++def chrome_path() -> str | None:+ if os.environ.get("CHROME"):+ return os.environ["CHROME"]+ for c in ("google-chrome", "chromium", "chromium-browser"):+ p = shutil.which(c)+ if p:+ return p+ return None+++def save_png(img, path: pathlib.Path) -> None:+ from PIL import Image+ q = img.convert("RGB").quantize(colors=96, method=Image.Quantize.MEDIANCUT, dither=Image.Dither.NONE)+ path.parent.mkdir(parents=True, exist_ok=True)+ q.save(path, optimize=True)+++def save_webp(img, path: pathlib.Path) -> None:+ """Lossless WebP of the same 96-colour image the PNG gets: flat inks and+ dot screens compress far better losslessly than with lossy WebP, and the+ dots stay crisp."""+ from PIL import Image+ q = img.convert("RGB").quantize(colors=96, method=Image.Quantize.MEDIANCUT, dither=Image.Dither.NONE)+ path.parent.mkdir(parents=True, exist_ok=True)+ q.convert("RGB").save(path, "WEBP", lossless=True, quality=100, method=6)+++def render(posts: list[dict], series_total: int, numbers: dict[str, int],+ html_dir: pathlib.Path | None) -> dict[str, dict]:+ from playwright.sync_api import sync_playwright+ from PIL import Image+ out: dict[str, dict] = {}+ with sync_playwright() as p:+ kw = {}+ exe = chrome_path()+ try:+ browser = p.chromium.launch()+ except Exception:+ if not exe:+ raise+ browser = p.chromium.launch(executable_path=exe)+ for post in posts:+ out[post["slug"]] = {}+ for kind, (W, H) in SIZES.items():+ pg = browser.new_page(viewport={"width": W, "height": H}, device_scale_factor=1)+ doc = page(post, kind, numbers[post["slug"]], series_total)+ if html_dir:+ html_dir.mkdir(parents=True, exist_ok=True)+ (html_dir / f"{post['slug']}-{kind}.html").write_text(doc)+ tmp = HERE / f".render-{kind}.html"+ tmp.write_text(doc)+ pg.goto(tmp.as_uri())+ pg.evaluate("document.fonts.ready")+ pg.wait_for_timeout(150)+ pg.evaluate(FIT_ART_JS if kind in ART_ONLY else FIT_JS)+ png = pg.locator(".card").screenshot(type="png")+ tmp.unlink()+ pg.close()+ out[post["slug"]][kind] = Image.open(io.BytesIO(png)).convert("RGB")+ browser.close()+ return out+++def framed(img):+ """The og card gets a 1px frame in TOKENS["frame"]: paper on a white+ timeline (X, Slack, LinkedIn) otherwise loses its edge. The hero and the+ thumbnail are framed by blog.css instead."""+ from PIL import ImageDraw+ out = img.copy()+ ImageDraw.Draw(out).rectangle([0, 0, out.width - 1, out.height - 1], outline=TOKENS["frame"])+ return out+++def write(images: dict[str, dict]) -> list[pathlib.Path]:+ from PIL import Image+ written = []+ for slug, imgs in images.items():+ og = LANDING / "og" / f"blog-{slug}.png"+ save_png(framed(imgs["og"]), og)+ hero = LANDING / "blog" / "covers" / f"{slug}-hero.webp"+ save_webp(imgs["hero"], hero)+ thumb = LANDING / "blog" / "covers" / f"{slug}-thumb.webp"+ # the thumbnail is a downscale, full of in-between tones: lossy wins there+ thumb.parent.mkdir(parents=True, exist_ok=True)+ imgs["thumb"].resize(THUMB, Image.LANCZOS).save(thumb, "WEBP", quality=80, method=6)+ written += [og, hero, thumb]+ return written+++def contact_sheet(images: dict[str, dict], path: pathlib.Path,+ cell: tuple[int, int] = (600, 315)) -> None:+ from PIL import Image+ cols = 3+ cw, ch = cell+ gap = 24+ rows = (len(images) + cols - 1) // cols+ sheet = Image.new("RGB", (cols * cw + (cols + 1) * gap, rows * ch + (rows + 1) * gap), (255, 255, 255))+ for n, (slug, imgs) in enumerate(images.items()):+ r, c = divmod(n, cols)+ sheet.paste(framed(imgs["og"].resize((cw, ch), Image.LANCZOS)), (gap + c * (cw + gap), gap + r * (ch + gap)))+ path.parent.mkdir(parents=True, exist_ok=True)+ sheet.save(path, optimize=True)+++def main() -> int:+ ap = argparse.ArgumentParser(description=__doc__.split("\n")[0])+ ap.add_argument("--only", action="append", help="render only this slug (repeatable)")+ ap.add_argument("--contact", type=pathlib.Path, help="also write a contact sheet PNG of every og card")+ ap.add_argument("--html-dir", type=pathlib.Path, help="keep the intermediate HTML pages here")+ a = ap.parse_args()+ numbers = {p["slug"]: i + 1 for i, p in enumerate(POSTS)}+ todo = [p for p in POSTS if not a.only or p["slug"] in a.only]+ if a.only and len(todo) != len(a.only):+ known = {p["slug"] for p in POSTS}+ print("unknown slug(s): " + ", ".join(s for s in a.only if s not in known), file=sys.stderr)+ return 2+ images = render(todo, len(POSTS), numbers, a.html_dir)+ for f in write(images):+ print(f"wrote {f.relative_to(REPO)} ({f.stat().st_size:,} B)")+ if a.contact:+ contact_sheet(images, a.contact)+ print(f"wrote {a.contact}")+ art = a.contact.with_name(a.contact.stem + "-art" + a.contact.suffix)+ contact_sheet({k: {"og": v["hero"]} for k, v in images.items()}, art, (600, 240))+ print(f"wrote {art}")+ return 0+++if __name__ == "__main__":+ sys.exit(main())diff --git a/scripts/seo/blog_covers/fonts/BigShouldersDisplay-wght.ttf b/scripts/seo/blog_covers/fonts/BigShouldersDisplay-wght.ttfnew file mode 100644index 000000000..abb8c8d40Binary files /dev/null and b/scripts/seo/blog_covers/fonts/BigShouldersDisplay-wght.ttf differdiff --git a/scripts/seo/blog_covers/fonts/JetBrainsMono-wght.ttf b/scripts/seo/blog_covers/fonts/JetBrainsMono-wght.ttfnew file mode 100644index 000000000..4c96e7985Binary files /dev/null and b/scripts/seo/blog_covers/fonts/JetBrainsMono-wght.ttf differdiff --git a/scripts/seo/blog_covers/fonts/OFL-BigShouldersDisplay.txt b/scripts/seo/blog_covers/fonts/OFL-BigShouldersDisplay.txtnew file mode 100644index 000000000..b5a4e0dff--- /dev/null+++ b/scripts/seo/blog_covers/fonts/OFL-BigShouldersDisplay.txt@@ -0,0 +1,93 @@+Copyright 2019 The Big Shoulders Project Authors (https://github.com/xotypeco/big_shoulders)++This Font Software is licensed under the SIL Open Font License, Version 1.1.+This license is copied below, and is also available with a FAQ at:+http://scripts.sil.org/OFL+++-----------------------------------------------------------+SIL OPEN FONT LICENSE Version 1.1 - 26 February 2007+-----------------------------------------------------------++PREAMBLE+The goals of the Open Font License (OFL) are to stimulate worldwide+development of collaborative font projects, to support the font creation+efforts of academic and linguistic communities, and to provide a free and+open framework in which fonts may be shared and improved in partnership+with others.++The OFL allows the licensed fonts to be used, studied, modified and+redistributed freely as long as they are not sold by themselves. The+fonts, including any derivative works, can be bundled, embedded,+redistributed and/or sold with any software provided that any reserved+names are not used by derivative works. The fonts and derivatives,+however, cannot be released under any other type of license. The+requirement for fonts to remain under this license does not apply+to any document created using the fonts or their derivatives.++DEFINITIONS+"Font Software" refers to the set of files released by the Copyright+Holder(s) under this license and clearly marked as such. This may+include source files, build scripts and documentation.++"Reserved Font Name" refers to any names specified as such after the+copyright statement(s).++"Original Version" refers to the collection of Font Software components as+distributed by the Copyright Holder(s).++"Modified Version" refers to any derivative made by adding to, deleting,+or substituting -- in part or in whole -- any of the components of the+Original Version, by changing formats or by porting the Font Software to a+new environment.++"Author" refers to any designer, engineer, programmer, technical+writer or other person who contributed to the Font Software.++PERMISSION & CONDITIONS+Permission is hereby granted, free of charge, to any person obtaining+a copy of the Font Software, to use, study, copy, merge, embed, modify,+redistribute, and sell modified and unmodified copies of the Font+Software, subject to the following conditions:++1) Neither the Font Software nor any of its individual components,+in Original or Modified Versions, may be sold by itself.++2) Original or Modified Versions of the Font Software may be bundled,+redistributed and/or sold with any software, provided that each copy+contains the above copyright notice and this license. These can be+included either as stand-alone text files, human-readable headers or+in the appropriate machine-readable metadata fields within text or+binary files as long as those fields can be easily viewed by the user.++3) No Modified Version of the Font Software may use the Reserved Font+Name(s) unless explicit written permission is granted by the corresponding+Copyright Holder. This restriction only applies to the primary font name as+presented to the users.++4) The name(s) of the Copyright Holder(s) or the Author(s) of the Font+Software shall not be used to promote, endorse or advertise any+Modified Version, except to acknowledge the contribution(s) of the+Copyright Holder(s) and the Author(s) or with their explicit written+permission.++5) The Font Software, modified or unmodified, in part or in whole,+must be distributed entirely under this license, and must not be+distributed under any other license. The requirement for fonts to+remain under this license does not apply to any document created+using the Font Software.++TERMINATION+This license becomes null and void if any of the above conditions are+not met.++DISCLAIMER+THE FONT SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTIES OF+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT+OF COPYRIGHT, PATENT, TRADEMARK, OR OTHER RIGHT. IN NO EVENT SHALL THE+COPYRIGHT HOLDER BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,+INCLUDING ANY GENERAL, SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL+DAMAGES, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING+FROM, OUT OF THE USE OR INABILITY TO USE THE FONT SOFTWARE OR FROM+OTHER DEALINGS IN THE FONT SOFTWARE.diff --git a/scripts/seo/blog_covers/fonts/OFL-JetBrainsMono.txt b/scripts/seo/blog_covers/fonts/OFL-JetBrainsMono.txtnew file mode 100644index 000000000..821a3dac2--- /dev/null+++ b/scripts/seo/blog_covers/fonts/OFL-JetBrainsMono.txt@@ -0,0 +1,93 @@+Copyright 2020 The JetBrains Mono Project Authors (https://github.com/JetBrains/JetBrainsMono)++This Font Software is licensed under the SIL Open Font License, Version 1.1.++This license is copied below, and is also available with a FAQ at: https://scripts.sil.org/OFL+++-----------------------------------------------------------+SIL OPEN FONT LICENSE Version 1.1 - 26 February 2007+-----------------------------------------------------------++PREAMBLE+The goals of the Open Font License (OFL) are to stimulate worldwide+development of collaborative font projects, to support the font creation+efforts of academic and linguistic communities, and to provide a free and+open framework in which fonts may be shared and improved in partnership+with others.++The OFL allows the licensed fonts to be used, studied, modified and+redistributed freely as long as they are not sold by themselves. The+fonts, including any derivative works, can be bundled, embedded,+redistributed and/or sold with any software provided that any reserved+names are not used by derivative works. The fonts and derivatives,+however, cannot be released under any other type of license. The+requirement for fonts to remain under this license does not apply+to any document created using the fonts or their derivatives.++DEFINITIONS+"Font Software" refers to the set of files released by the Copyright+Holder(s) under this license and clearly marked as such. This may+include source files, build scripts and documentation.++"Reserved Font Name" refers to any names specified as such after the+copyright statement(s).++"Original Version" refers to the collection of Font Software components as+distributed by the Copyright Holder(s).++"Modified Version" refers to any derivative made by adding to, deleting,+or substituting -- in part or in whole -- any of the components of the+Original Version, by changing formats or by porting the Font Software to a+new environment.++"Author" refers to any designer, engineer, programmer, technical+writer or other person who contributed to the Font Software.++PERMISSION & CONDITIONS+Permission is hereby granted, free of charge, to any person obtaining+a copy of the Font Software, to use, study, copy, merge, embed, modify,+redistribute, and sell modified and unmodified copies of the Font+Software, subject to the following conditions:++1) Neither the Font Software nor any of its individual components,+in Original or Modified Versions, may be sold by itself.++2) Original or Modified Versions of the Font Software may be bundled,+redistributed and/or sold with any software, provided that each copy+contains the above copyright notice and this license. These can be+included either as stand-alone text files, human-readable headers or+in the appropriate machine-readable metadata fields within text or+binary files as long as those fields can be easily viewed by the user.++3) No Modified Version of the Font Software may use the Reserved Font+Name(s) unless explicit written permission is granted by the corresponding+Copyright Holder. This restriction only applies to the primary font name as+presented to the users.++4) The name(s) of the Copyright Holder(s) or the Author(s) of the Font+Software shall not be used to promote, endorse or advertise any+Modified Version, except to acknowledge the contribution(s) of the+Copyright Holder(s) and the Author(s) or with their explicit written+permission.++5) The Font Software, modified or unmodified, in part or in whole,+must be distributed entirely under this license, and must not be+distributed under any other license. The requirement for fonts to+remain under this license does not apply to any document created+using the Font Software.++TERMINATION+This license becomes null and void if any of the above conditions are+not met.++DISCLAIMER+THE FONT SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTIES OF+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT+OF COPYRIGHT, PATENT, TRADEMARK, OR OTHER RIGHT. IN NO EVENT SHALL THE+COPYRIGHT HOLDER BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,+INCLUDING ANY GENERAL, SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL+DAMAGES, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING+FROM, OUT OF THE USE OR INABILITY TO USE THE FONT SOFTWARE OR FROM+OTHER DEALINGS IN THE FONT SOFTWARE.diff --git a/scripts/seo/blog_covers/motifs.py b/scripts/seo/blog_covers/motifs.pynew file mode 100644index 000000000..b55dd12cc--- /dev/null+++ b/scripts/seo/blog_covers/motifs.py@@ -0,0 +1,501 @@+"""Square-matrix motifs for the blog cover series ("engineering field-log posters").++Every motif is a pure function ``motif(w, h, rng) -> str`` that returns SVG+markup drawn inside a ``w`` x ``h`` box whose origin is the top-left corner.+They draw with squares on a grid only (plus 1px hairlines and mono labels);+``dot`` is a filled cell, ``ring`` an empty one. Inks, on light paper:++ .y yellow #ffc400 fill only: XERJ / the one result the post is about+ (its labels and hairlines print in ochre #a06800)+ .w ink #0b0b0d structure, neutral series+ .b blue #6f8aa8 the comparison series, never XERJ+ .m mute #696762 annotation+ .f rule #cfcbbf the empty cell, the background grid++Every number a motif draws or labels is quoted from the post it illustrates;+the comment above each motif names where. Randomness only ever decides+*placement* (jitter, scatter) and comes from ``rng``, a ``random.Random``+seeded with the post slug, so the output is identical on every run.+"""++from __future__ import annotations++import math+import random+from typing import Callable++# ── primitives ───────────────────────────────────────────────────────────────+++def _f(v: float) -> str:+ return f"{v:.1f}".rstrip("0").rstrip(".")+++def dot(x: float, y: float, r: float, ink: str) -> str:+ """One cell of the matrix: a square centred on (x, y). ``r`` is the+ half-pitch-scaled size the halftone asks for; the square keeps the same+ centre, so every screen stays on its grid."""+ if r <= 0.15:+ return ""+ s = r * 1.78+ return (f'<rect x="{_f(x - s / 2)}" y="{_f(y - s / 2)}" width="{_f(s)}" height="{_f(s)}" '+ f'class="{ink}"/>')+++def ring(x: float, y: float, r: float, ink: str, sw: float = 1.2) -> str:+ """An empty cell: a square outline, the matrix's "not this one"."""+ s = r * 1.78+ return (f'<rect x="{_f(x - s / 2)}" y="{_f(y - s / 2)}" width="{_f(s)}" height="{_f(s)}" '+ f'class="{ink}-s" fill="none" stroke-width="{_f(sw)}"/>')+++# Art-only renders (the in-post hero and the index thumbnail) set this to+# False: the page already carries the title, so those covers draw no text at+# all. The og card keeps its labels.+LABELS = True+++def label(x: float, y: float, text: str, cls: str = "lab", anchor: str = "start") -> str:+ if not LABELS:+ return ""+ text = (text.replace("&", "&").replace("<", "<").replace(">", ">"))+ return f'<text x="{_f(x)}" y="{_f(y)}" class="{cls}" text-anchor="{anchor}">{text}</text>'+++def hair(x1: float, y1: float, x2: float, y2: float, ink: str = "f", sw: float = 1) -> str:+ return (f'<line x1="{_f(x1)}" y1="{_f(y1)}" x2="{_f(x2)}" y2="{_f(y2)}" '+ f'class="{ink}-s" stroke-width="{_f(sw)}"/>')+++def halftone(x: float, y: float, w: float, h: float, pitch: float,+ val: Callable[[float, float], float], ink: str,+ rng: random.Random | None = None, grain: float = 0.0) -> str:+ """A dot screen over a rectangle; ``val(u, v)`` in 0..1 sets dot area at the+ normalised position (u, v). ``grain`` drops dots at random, risograph-like."""+ out = []+ cols = max(1, int(w // pitch))+ rows = max(1, int(h // pitch))+ rmax = pitch * 0.5+ for j in range(rows):+ for i in range(cols):+ u = (i + 0.5) / cols+ v = (j + 0.5) / rows+ a = max(0.0, min(1.0, val(u, v)))+ if rng is not None and grain and rng.random() < grain:+ continue+ r = rmax * math.sqrt(a) * 0.96+ if rng is not None:+ r *= 1 + rng.uniform(-0.07, 0.07)+ out.append(dot(x + (i + 0.5) * pitch, y + (j + 0.5) * pitch, r, ink))+ return "".join(out)+++def dotbar(x: float, y: float, length: float, rows: int, pitch: float, ink: str,+ fade: int = 5) -> str:+ """A horizontal bar made of a dot screen; the last ``fade`` columns thin out+ like an ink edge. ``length`` is honoured to the column."""+ cols = max(1, int(round(length / pitch)))+ out = []+ for i in range(cols):+ a = 1.0 if i < cols - fade else max(0.18, (cols - i) / (fade + 1))+ for j in range(rows):+ out.append(dot(x + (i + 0.5) * pitch, y + (j + 0.5) * pitch,+ pitch * 0.5 * math.sqrt(a) * 0.96, ink))+ return "".join(out)+++def dotpath(points: list[tuple[float, float]], step: float, r0: float, r1: float,+ ink: str) -> str:+ """Dots spaced ``step`` apart along a polyline, radius ramping r0 -> r1."""+ segs = list(zip(points, points[1:]))+ total = sum(math.dist(a, b) for a, b in segs) or 1.0+ out, walked, carry = [], 0.0, 0.0+ for a, b in segs:+ L = math.dist(a, b)+ t = carry+ while t <= L:+ k = (walked + t) / total+ px = a[0] + (b[0] - a[0]) * t / L+ py = a[1] + (b[1] - a[1]) * t / L+ out.append(dot(px, py, r0 + (r1 - r0) * k, ink))+ t += step+ carry = t - L+ walked += L+ return "".join(out)+++def bezier(p0, p1, p2, n: int = 60) -> list[tuple[float, float]]:+ return [((1 - t) ** 2 * p0[0] + 2 * (1 - t) * t * p1[0] + t * t * p2[0],+ (1 - t) ** 2 * p0[1] + 2 * (1 - t) * t * p1[1] + t * t * p2[1])+ for t in (i / n for i in range(n + 1))]+++def disc(cx: float, cy: float, R: float, pitch: float, ink: str,+ profile: Callable[[float], float] = lambda d: 1 - d * d) -> str:+ """A round halftone blob: dot area falls off with distance from the centre."""+ out = []+ n = int(R // pitch) + 1+ for j in range(-n, n + 1):+ for i in range(-n, n + 1):+ x, y = cx + i * pitch, cy + j * pitch+ d = math.hypot(x - cx, y - cy) / R+ if d > 1:+ continue+ out.append(dot(x, y, pitch * 0.5 * math.sqrt(max(0, profile(d))) * 0.96, ink))+ return "".join(out)+++# ── motifs ───────────────────────────────────────────────────────────────────++def m_index28(w: float, h: float, rng: random.Random) -> str:+ """cutting-the-index-28-percent: the index as a block of squares, before and+ after. Areas are to scale: 5.52 MB before, 3.99 MB after (-28%), the whole-+ index figures in the post's lede and index card. The cut-away (the 28%)+ is the yellow L. The two side bars are the post's .seg (1,671 -> 449 kB,+ -73.1%) and .meta (200.8 -> 13.8 kB, -93.1%) rows, each against its own+ 'before' length."""+ out = []+ S = min(h * 0.80, w * 0.58)+ pitch = max(8.0, S / 26)+ n = int(S // pitch)+ S = n * pitch+ x0, y0 = 0.0, 26.0+ k = int(round(n * math.sqrt(3.99 / 5.52))) # after-block side, in cells+ for j in range(n):+ for i in range(n):+ cx, cy = x0 + (i + 0.5) * pitch, y0 + (j + 0.5) * pitch+ if i < k and j >= n - k:+ u, v = i / k, (j - (n - k)) / k+ out.append(dot(cx, cy, pitch * 0.40, "w"))+ else:+ out.append(dot(cx, cy, pitch * 0.40, "y"))+ out.append(label(x0, y0 - 12, "BEFORE · 5.52 MB", "lab"))+ out.append(label(x0, y0 + S + 22, "AFTER · 3.99 MB", "lab w"))+ bx = x0 + S + 30+ bw = w - bx+ out.append(label(bx, y0 + 40, "−28%", "big w"))+ if bw > 110:+ rows = [(".SEG", 1671, 449, "−73.1%"), (".META", 200.8, 13.8, "−93.1%")]+ by = y0 + S * 0.38+ bp = max(5.0, pitch * 0.55)+ for name, before, after, delta in rows:+ out.append(label(bx, by, f"{name} {delta}", "lab"))+ out.append(dotbar(bx, by + 10, bw, 2, bp, "f", 3))+ out.append(dotbar(bx, by + 14 + 2 * bp, max(bp, bw * after / before), 2, bp, "w", 1))+ by += S * 0.30+ return "".join(out)+++def m_jevbars(w: float, h: float, rng: random.Random) -> str:+ """does-xerj-beat-jev: the verdict table as square-bars. nDCG@10 from the post:+ SciFact 0.6993 (hybrid) vs 0.7410 (BM25+Jev), NFCorpus 0.3448 vs 0.3312,+ FiQA 0.2382 (BM25) vs 0.3638 (BM25+Jev). One denominator, 0.7410, for every+ row. Last row: the FiQA rerank bill, $0.6375 vs $0 local."""+ out = []+ rows = [("SCIFACT", 0.6993, "hybrid", 0.7410, "BM25+Jev"),+ ("NFCORPUS", 0.3448, "hybrid", 0.3312, "BM25+Jev"),+ ("FIQA", 0.2382, "BM25", 0.3638, "BM25+Jev")]+ lab_w = 0.0+ full = w * 0.70+ gap = h / 4.6+ pitch = max(6.0, gap / 9.5)+ y = 8.0+ for name, a, an, b, bn in rows:+ out.append(label(0, y + 12, name, "lab"))+ out.append(dotbar(lab_w, y + 22, full * a / 0.7410, 3, pitch, "w"))+ out.append(label(lab_w + full * a / 0.7410 + 10, y + 22 + 2.2 * pitch, f"{a:.4f} {an}", "num w"))+ out.append(dotbar(lab_w, y + 26 + 3 * pitch, full * b / 0.7410, 3, pitch, "b"))+ out.append(label(lab_w + full * b / 0.7410 + 10, y + 26 + 5.2 * pitch, f"{b:.4f} {bn}", "num b"))+ y += gap+ out.append(hair(0, y - 4, w, y - 4, "f"))+ out.append(label(0, y + 16, "FIQA RERANK BILL", "lab"))+ out.append(dotbar(lab_w, y + 26, full, 3, pitch, "b"))+ out.append(label(lab_w + full + 10, y + 26 + 2.2 * pitch, "$0.6375", "num b"))+ out.append(f'<rect x="{_f(lab_w)}" y="{_f(y + 30 + 3.4 * pitch)}" width="{_f(2.2 * pitch)}" height="{_f(2.2 * pitch)}" class="yo"/>')+ out.append(label(lab_w + 2.2 * pitch + 10, y + 30 + 5.0 * pitch, "$0 · XERJ local", "num y"))+ return "".join(out)+++def m_vote(w: float, h: float, rng: random.Random) -> str:+ """jev-vs-a-bm25-vote: Banking77, 500 decisions per arm as a 25x20+ square grid; filled = correct, outline = wrong (yellow: the vote's 45+ correct answers beyond Jev's 412), at the post's accuracies 0.9140 (BM25+ vote) and 0.8240 (Jev). The footer line is the SMS row: 0.9840 vs 0.9880,+ McNemar p = 0.75."""+ out = []+ cols, rows_n = 25, 20+ gw = (w - 36) / 2+ pitch = min(gw / cols, (h - 92) / rows_n)+ gy = 34.0+ jev_right = round(0.8240 * 500)+ for k, (name, acc, ink) in enumerate([("BM25 VOTE", 0.9140, "w"), ("JEV", 0.8240, "b")]):+ gx = k * (gw + 36)+ right = round(acc * 500)+ out.append(label(gx, gy - 14, f"{name} · {acc:.4f}", f"lab {ink}"))+ for n in range(500):+ i, j = n % cols, n // cols+ cx, cy = gx + (i + 0.5) * pitch, gy + (j + 0.5) * pitch+ if n < right:+ # yellow: the decisions the vote gets right that Jev does not+ out.append(dot(cx, cy, pitch * 0.40, "y" if (k == 0 and n >= jev_right) else ink))+ else:+ out.append(ring(cx, cy, pitch * 0.30, "m", 1.0))+ yb = gy + rows_n * pitch + 26+ out.append(label(0, yb, "BANKING77 · 77 CLASSES · 500 DECISIONS EACH", "lab"))+ out.append(label(0, yb + 22, "SMS SPAM 0.9840 vs 0.9880 · McNEMAR p = 0.75 · A TIE", "lab m"))+ return "".join(out)+++def m_names(w: float, h: float, rng: random.Random) -> str:+ """one-vulnerability-three-names: RUSTSEC-2021-0003, GHSA-43w2-9j62-hq99 and+ CVE-2021-25900 (the post's lede) converging on one identity-resolved+ record; caption 4,107 filings -> 1,950 records, the post's own count."""+ out = []+ names = ["RUSTSEC-2021-0003", "GHSA-43w2-9j62-hq99", "CVE-2021-25900"]+ R = min(h * 0.27, w * 0.2)+ cx, cy = w - R - 6, h * 0.46+ out.append(disc(cx, cy, R, max(7.0, R / 9), "w"))+ c = max(7.0, R / 9) * 1.6+ out.append(f'<rect x="{_f(cx - c)}" y="{_f(cy - c)}" width="{_f(2 * c)}" height="{_f(2 * c)}" class="yo"/>')+ out.append(ring(cx, cy, R + 12, "y", 1.0))+ ys = [h * 0.12, h * 0.46, h * 0.80]+ for n, y in zip(names, ys):+ out.append(f'<rect x="0" y="{_f(y - 17)}" width="{_f(w * 0.40)}" height="30" class="w-s" fill="none" stroke-width="1"/>')+ out.append(label(10, y + 3, n, "num w"))+ p0 = (w * 0.40 + 6, y - 2)+ p2 = (cx - R - 16, cy)+ p1 = ((p0[0] + p2[0]) / 2, y)+ out.append(dotpath(bezier(p0, p1, p2), 9, 1.2, 3.2, "w"))+ out.append(label(cx, cy + R + 40, "1 RECORD · ed25519-SIGNED", "lab y", "middle"))+ out.append(label(0, h - 2, "4,107 FILINGS → 1,950 RECORDS", "lab"))+ return "".join(out)+++def _gate(x: float, y: float, gw: float, gh: float, pitch: float, post_ink: str,+ glyph: str, glyph_ink: str) -> str:+ out = []+ cols = 2+ for side in (0, 1):+ px = x + (0 if side == 0 else gw - cols * pitch)+ ink = post_ink if not (post_ink == "split") else ("y" if side == 0 else "w")+ out.append(halftone(px, y, cols * pitch, gh, pitch, lambda u, v: 1.0, ink))+ lint_ink = "y" if post_ink == "split" else post_ink+ out.append(halftone(x, y - pitch * 1.4, gw, pitch, pitch, lambda u, v: 1.0, lint_ink))+ mx, my = x + gw / 2, y + gh * 0.45+ s = gw * 0.20+ if glyph == "check":+ out.append(dotpath([(mx - s, my), (mx - s * 0.25, my + s * 0.8), (mx + s * 1.1, my - s)], pitch * 0.8, 2.6, 2.6, glyph_ink))+ elif glyph == "cross":+ out.append(dotpath([(mx - s, my - s), (mx + s, my + s)], pitch * 0.8, 2.6, 2.6, glyph_ink))+ out.append(dotpath([(mx + s, my - s), (mx - s, my + s)], pitch * 0.8, 2.6, 2.6, glyph_ink))+ elif glyph == "half":+ out.append(dotpath([(mx - s, my), (mx - s * 0.25, my + s * 0.8), (mx + s * 0.2, my + s * 0.2)], pitch * 0.8, 2.6, 2.6, "y"))+ out.append(dotpath([(mx + s * 0.35, my - s), (mx + s * 1.1, my - s * 0.2)], pitch * 0.8, 2.6, 2.6, "w"))+ elif glyph == "dash":+ out.append(dotpath([(mx - s * 0.7, my), (mx + s * 0.7, my)], pitch * 0.8, 2.2, 2.2, glyph_ink))+ return "".join(out)+++def m_gates(w: float, h: float, rng: random.Random) -> str:+ """the-rc80-gates: the program table, one gate per feature, in the post's+ order of verdicts: 3 PASS (xerj_map, /_ask, p_cal), 3 FAIL (decide head,+ judge, flywheel), 1 SPLIT (stemming), 1 REPORTED (the open model card), and+ the two rows the post marks unmeasured on purpose (max_tokens, _watcher)."""+ gates = [("xerj_map", "PASS"), ("/_ask", "PASS"), ("p_cal", "PASS"),+ ("decide", "FAIL"), ("judge", "FAIL"), ("flywheel", "FAIL"),+ ("stemming", "SPLIT"), ("model card", "REPORTED"),+ ("max_tokens", "NO GATE"), ("_watcher", "NOT RUN")]+ style = {"PASS": ("y", "check", "y"), "FAIL": ("w", "cross", "w"),+ "SPLIT": ("split", "half", "y"), "REPORTED": ("m", "dash", "m"),+ "NO GATE": ("f", "dash", "f"), "NOT RUN": ("f", "dash", "f")}+ out = []+ per_row = 5+ cw = w / per_row+ gw = cw * 0.62+ rh = (h - 40) / 2+ gh = rh * 0.52+ pitch = max(5.0, gw / 11)+ for n, (feat, verdict) in enumerate(gates):+ r, c = divmod(n, per_row)+ x = c * cw + (cw - gw) / 2 - cw * 0.06+ y = r * rh + rh * 0.18 + 12+ post, glyph, gink = style[verdict]+ out.append(_gate(x, y, gw, gh, pitch, post, glyph, gink))+ vcls = {"PASS": "lab y", "FAIL": "lab w", "SPLIT": "lab y"}.get(verdict, "lab m")+ out.append(label(x, y + gh + 22, verdict, vcls))+ out.append(label(x, y + gh + 40, feat, "num m"))+ out.append(label(0, h - 2, "10 FEATURES · 8 GATES · 3 PASS · 3 FAIL · 1 SPLIT", "lab"))+ return "".join(out)+++def m_tiles(w: float, h: float, rng: random.Random) -> str:+ """the-corpus-hub: the registry as a 10x10 wall of pack tiles, 100 being+ the live-corpus count in the post's 2026-10-06 update. The four launch+ corpora are solid yellow; the one signed daily pack carries a seal frame."""+ out = []+ n = 10+ side = min((h - 44) / n, w * 0.78 / n)+ gap = side * 0.16+ t = side - gap+ sub = 4+ pitch = t / sub+ x0 = 0.0+ y0 = 6.0+ for k in range(100):+ i, j = k % n, k // n+ x, y = x0 + i * side, y0 + j * side+ if k < 4:+ out.append(halftone(x, y, t, t, pitch, lambda u, v: 1.0, "y"))+ else:+ d = 0.10 + 0.55 * rng.random()+ out.append(halftone(x, y, t, t, pitch, lambda u, v, d=d: d, "w"))+ sx, sy = x0 + 4 * side + t / 2, y0 + t / 2+ out.append(ring(sx, sy, t * 0.62, "y", 2))+ lx = x0 + n * side + 18+ if lx < w - 80:+ out.append(label(lx, y0 + 14, "4 AT LAUNCH", "lab y"))+ out.append(label(lx, y0 + 34, "+1 SIGNED PACK", "lab y"))+ out.append(label(lx, y0 + n * side - 26, "100 LIVE", "lab"))+ out.append(label(lx, y0 + n * side - 6, "2026-10-06", "lab m"))+ out.append(label(0, h - 2, "PINNED · LICENCED · CHECKSUMMED · SIGNED", "lab"))+ return "".join(out)+++def m_network(w: float, h: float, rng: random.Random) -> str:+ """the-exploit-hub: a field of PoC documents and the three fresh CVEs of the+ eval. CVE-2026-23744 links to 36 PoC nodes (the post's 36 in-corpus PoC+ repositories); CVE-2026-105844 had no PoC anywhere, so one link to a+ neighbouring 2026 exploit; CVE-2026-97332 is the empty stratum, unlinked."""+ out = []+ H = h - 30+ pts = []+ tries = 0+ while len(pts) < 150 and tries < 6000:+ tries += 1+ p = (rng.uniform(8, w - 8), rng.uniform(8, H - 8))+ if all(math.dist(p, q) > 22 for q in pts):+ pts.append(p)+ for p in pts:+ near = sorted(pts, key=lambda q: math.dist(p, q))[1:3]+ for q in near:+ out.append(hair(p[0], p[1], q[0], q[1], "f", 0.8))+ A = (w * 0.30, H * 0.46)+ B = (w * 0.78, H * 0.22)+ C = (w * 0.78, H * 0.78)+ near_a = sorted(pts, key=lambda q: math.dist(A, q))[:36]+ for q in near_a:+ out.append(hair(A[0], A[1], q[0], q[1], "y", 0.9))+ for p in pts:+ hot = p in near_a+ out.append(dot(p[0], p[1], 3.2 if hot else 2.2, "y" if hot else "m"))+ nb = min((q for q in pts if math.dist(q, B) > 60), key=lambda q: math.dist(q, B))+ out.append(dotpath([B, nb], 8, 2.0, 2.0, "w"))+ out.append(dot(nb[0], nb[1], 4.5, "w"))+ out.append(disc(A[0], A[1], 30, 6.5, "y"))+ out.append(disc(B[0], B[1], 20, 6, "w"))+ out.append(f'<circle cx="{_f(C[0])}" cy="{_f(C[1])}" r="40" class="bgfill"/>')+ out.append(ring(C[0], C[1], 18, "w", 1.6))+ out.append(label(A[0], A[1] + 52, "CVE-2026-23744 · 36 POCS", "tag y", "middle"))+ out.append(label(B[0], B[1] + 40, "CVE-2026-105844 · A NEIGHBOUR", "tag w", "middle"))+ out.append(label(C[0], C[1] + 38, "CVE-2026-97332 · ADDS ~0", "tag m", "middle"))+ out.append(label(0, h - 2, "1.09M POC DOCUMENTS · 3 FRESH CVES · LOSSES COUNTED", "lab"))+ return "".join(out)+++def m_scars(w: float, h: float, rng: random.Random) -> str:+ """eval-scars-from-agent-memory: six scratch marks, one per pattern the+ guest post credits, each tagged with that pattern's own figure from the+ post: topic slicing at recall@1, the +2.75pt incumbent, group-by-qid, the+ 0.87 classifier vs the 0.78 always-yes baseline, the ~4x unpaired+ overstatement, and the mask bug that scores every query zero."""+ tags = ["TOPIC SLICING · r@1", "+2.75pt INCUMBENT", "GROUP BY QID",+ "0.87 vs 0.78 ALWAYS-YES", "4× UNPAIRED", "MASK → 0 HITS"]+ out = []+ H = h - 30+ # faint scatter of nicks across the whole field+ for _ in range(90):+ x, y = rng.uniform(0, w), rng.uniform(0, H)+ a = rng.uniform(-1.2, -0.6)+ L = rng.uniform(6, 16)+ out.append(dotpath([(x, y), (x + L * math.cos(a), y + L * math.sin(a))], 4, 1.0, 1.0, "f"))+ cw, rh = w / 3, H / 2+ for k, tag in enumerate(tags):+ r, c = divmod(k, 3)+ cx = c * cw + cw * rng.uniform(0.30, 0.55)+ cy = r * rh + rh * rng.uniform(0.32, 0.50)+ ink = "y" if k == 5 else "w"+ a = rng.uniform(-1.25, -0.85)+ strokes = 3 + (k % 3)+ L = rh * 0.52+ for s in range(strokes):+ ox = cx + (s - strokes / 2) * 11 + rng.uniform(-2, 2)+ oy = cy + rng.uniform(-4, 4)+ p0 = (ox - L / 2 * math.cos(a), oy - L / 2 * math.sin(a))+ p1 = (ox + L / 2 * math.cos(a), oy + L / 2 * math.sin(a))+ out.append(dotpath([p0, p1], 5.2, 1.4, 3.0, ink))+ out.append(label(c * cw + 4, r * rh + rh * 0.92, f"{k + 1:02d} {tag}", f"tag {ink}"))+ out.append(label(0, h - 2, "SIX PATTERNS · EACH COST A REAL FAILURE", "lab"))+ return "".join(out)+++def m_route(w: float, h: float, rng: random.Random) -> str:+ """corpus-hub-v2: the router as a wall of pack tiles. Back wall: 251 faint+ tiles, the post's count of distinct slugs with a signed v2 pack on GitHub+ Releases. Front: the 14 corpora in the signed catalogue the router reads+ (generated 2026-10-18), by slug. The question from the post's lede routes+ to lz4, whose 4.4 MB pack is pulled to a local node on port 9300. Timings+ are the post's table: 1.72-2.27 s cold, 37-57 ms warm."""+ slugs = ["act", "age", "agentops", "aiohttp", "alertmanager", "arkworks-algebra",+ "arroyo", "astro", "attack-mappings-explorer", "autorag", "avro",+ "lz4", "snappy", "zstd"]+ out = []+ # back wall: 251 packs on Releases (re-counted 2026-10-20); 36 x 7 holds them+ cols = 36+ bp = w / cols+ for k in range(251):+ i, j = k % cols, k // cols+ out.append(halftone(i * bp + 1, 30 + j * bp + 1, bp - 3, bp - 3, (bp - 3) / 2,+ lambda u, v: 0.5, "f"))+ wall_h = 30 + 7 * bp+ out.append(label(0, 18, "xerj \"how does lz4 decompress a block\"", "num w"))+ # front: the 14 catalogue tiles+ tc = 7+ tw = (w - 6 * 10) / tc+ ty0 = wall_h + 22+ # two tile rows, then the node and the caption must fit under them+ th = min(tw * 0.62, (h - 40 - ty0 - 34) / 2 - 30)+ lz = None+ for k, slug in enumerate(slugs):+ i, j = k % tc, k // tc+ x, y = i * (tw + 10), ty0 + j * (th + 30)+ hot = slug == "lz4"+ out.append(halftone(x, y, tw, th, tw / 9, (lambda u, v: 0.35 + 0.65 * (1 - v)) if hot+ else (lambda u, v: 0.16), "y" if hot else "w", rng))+ name = slug if len(slug) <= 9 else slug[:8] + "…"+ out.append(label(x, y + th + 15, name, "tag y" if hot else "tag m"))+ if hot:+ lz = (x + tw / 2, y + th / 2)+ # question -> lz4 -> node+ gx = 4 * (tw + 10) - 5 # the gutter left of lz4's column+ out.append(dotpath([(8, 26), (8, ty0 - 11), (gx, ty0 - 11), (gx, lz[1]),+ (lz[0] - tw / 2 - 3, lz[1])], 7, 1.6, 2.8, "y"))+ ny = ty0 + 2 * (th + 30) + 22+ out.append(dotpath([(lz[0], lz[1] + th / 2 + 22), (lz[0], ny - 14)], 7, 2.6, 2.6, "y"))+ out.append(ring(lz[0], ny, 12, "y", 2))+ out.append(dot(lz[0], ny, 5, "y"))+ out.append(label(lz[0] - 22, ny + 5, "4.4 MB PACK → LOCAL NODE :9300", "lab y", "end"))+ out.append(label(0, h - 2, "14 IN THE SIGNED CATALOGUE · 1.72–2.27 s COLD · 37–57 ms WARM", "lab"))+ return "".join(out)+++MOTIFS: dict[str, Callable[[float, float, random.Random], str]] = {+ "index28": m_index28,+ "jevbars": m_jevbars,+ "vote": m_vote,+ "names": m_names,+ "gates": m_gates,+ "tiles": m_tiles,+ "network": m_network,+ "scars": m_scars,+ "route": m_route,+}diff --git a/scripts/seo/pagedata.py b/scripts/seo/pagedata.pyindex 3d58f2c97..6e90e74c4 100755--- a/scripts/seo/pagedata.py+++ b/scripts/seo/pagedata.py@@ -24,7 +24,10 @@ Fieldsog_image optional per-page Open Graph card, as a landing-relative PNGpath under og/ (e.g. "og/does-xerj-beat-jev.png"). Omit tokeep the site-wide card. og_image_alt accompanies it; both- width and height stay the site-wide 1200x630.+ width and height stay the site-wide 1200x630. Blog posts use+ og/blog-<slug>.png, rendered with the in-post hero and the+ index thumbnail by scripts/seo/blog_covers/covers.py; keep+ og_image_alt equal to that post's `og_alt` in covers.POSTS.published / optional pinned dates for article-kind pages (the blog).updated When set they feed datePublished/dateModified verbatim, sothe generated head does not drift with git history.@@ -88,6 +91,9 @@ PAGES: dict[str, dict[str, str]] = {description="XERJ v1.0.0-rc.6 against Elasticsearch 8.13.4 across 88 measured cells — 55 win, 26 tie, 4 lose, 3 n/a — plus the four commands that reproduce every one.",),# ── blog ────────────────────────────────────────────────────────────────+ # Every post's og_image is its cover from the field-log series+ # (scripts/seo/blog_covers/, README there); og_image_alt describes only+ # what that card draws.# The Jev write-up lives under /blog since 2026-09-21; it was git-mv'd# from benchmarks/decisions.html (which 301s to it via landing/_redirects).# published/updated are pinned so datePublished/dateModified do not drift@@ -100,12 +106,8 @@ PAGES: dict[str, dict[str, str]] = {# 2026-09-27: added the closing 05 section (verdict + try-it); measurement# content unchanged. updated follows the edit, published stays pinned.updated="2026-09-27",- og_image="og/does-xerj-beat-jev.png",- # alt text must describe only what the committed 1200x630 card- # renders — the token count is NOT on the image (strip: queries,- # $0.6375 vs $0, URL); scripts/seo/assets/og-blog-*.html is the- # source of record for what is on it.- og_image_alt="XERJ vs JEV, measured: SciFact, NFCorpus and FiQA nDCG@10 over 1,271 judged queries, and the FiQA judge bill — $0.6375 against $0 local",+ og_image="og/blog-does-xerj-beat-jev.png",+ og_image_alt="Cover: “Does XERJ beat JEV?” in display type beside the post’s square-matrix art, labelled with its figures. The art: three pairs of square-matrix bars, ink for XERJ and blue for BM25+Jev, over one long blue bar and a single yellow square for the zero-cost local run",),# Decisions study, 2026-09-22. published/updated pinned so the schema dates# do not drift with git (same reason as the post above). Prose written to the@@ -118,6 +120,8 @@ PAGES: dict[str, dict[str, str]] = {# 2026-09-27: added the closing 08 section (takeaway + try-it); measurement# content unchanged. updated follows the edit, published stays pinned.updated="2026-09-27",+ og_image="og/blog-jev-vs-a-bm25-vote.png",+ og_image_alt="Cover: “A BM25 vote ties Jev on spam and beats it on intent” in display type beside the post’s square-matrix art, labelled with its figures. The art: two grids of 500 squares, filled for correct answers; the ink grid has a band of yellow squares where it beats the blue grid",),# Size-epic stage-1 write-up, 2026-09-27. published/updated pinned like the# two posts above; in-body run dates stay 2026-09-26/27.@@ -126,6 +130,8 @@ PAGES: dict[str, dict[str, str]] = {title="Cutting the index 28%: dissect first, write Rust second",description="Four file-format changes took our 100k-doc index from 5.52 MB to 3.99 MB (−28%), each gated on dissecting real segments first. The losing rows are in the post.",published="2026-09-27", updated="2026-09-27",+ og_image="og/blog-cutting-the-index-28-percent.png",+ og_image_alt="Cover: “Cutting the index 28%” in display type beside the post’s square-matrix art, labelled with its figures. The art: a large ink block of squares wrapped in a yellow L-shaped band that marks the 28% cut, beside two pairs of thin before-and-after bars",),# Corpus-pack GTM post, 2026-09-27. Every count traces to tools/packs/# rust-vulns/pack-stats.json (provenance "measured", pinned by test); the@@ -135,6 +141,8 @@ PAGES: dict[str, dict[str, str]] = {title="One vulnerability, three names — and a pack we sign",description="The same Rust vulnerability arrives as RUSTSEC, GHSA and CVE. We collapsed 4,107 advisories to 1,950 identity-resolved records, packed them, and signed the pack — because no vulnerability database we surveyed does.",published="2026-09-27", updated="2026-09-27",+ og_image="og/blog-one-vulnerability-three-names.png",+ og_image_alt="Cover: “One vulnerability, three names” in display type beside the post’s square-matrix art, labelled with its figures. The art: three outlined boxes trailing lines of small squares that converge on one square-matrix disc with a yellow seal at its centre",),# The rc.80 gates write-up, 2026-09-30 — the blog half of the packaging# rule for the ten-item program (#1055-#1064): every gate verdict, wins@@ -145,6 +153,8 @@ PAGES: dict[str, dict[str, str]] = {title="The rc.80 gates: three passed outright, three failed, one split",description="Ten features, eight measured gates: /_ask macro-F1 0.9975, xerj_map 20/30 to 0/30, calibration ECE 0.0088 pass; the judge, the flywheel and the decide head fail their bars, numbers published.",published="2026-09-30", updated="2026-09-30",+ og_image="og/blog-the-rc80-gates.png",+ og_image_alt="Cover: “The rc.80 gates” in display type beside the post’s square-matrix art, labelled with its figures. The art: ten gates built from squares in two rows: three yellow with check marks, three ink with crosses, one half yellow and three faint",),# Corpus Hub launch post, 2026-10-01. Numbers trace to: docs/case-studies/# reference-coding/SUMMARY.md (16/16, 9,982 vs 26,477, $1.58 vs $3.27),@@ -156,6 +166,8 @@ PAGES: dict[str, dict[str, str]] = {title="The Corpus Hub: pinned, licenced, signed, contributed by PR",description="A public registry of corpora for AI agents: reference code pinned to SHAs, record packs ed25519-signed, every source licence-reviewed by a human. The measured use case (2.7× fewer output tokens), the tie we published anyway, and how to contribute yours by PR.",published="2026-10-01", updated="2026-10-01",+ og_image="og/blog-the-corpus-hub.png",+ og_image_alt="Cover: “The Corpus Hub” in display type beside the post’s square-matrix art, labelled with its figures. The art: a 10 by 10 wall of square-matrix pack tiles in shades of ink, the first four solid yellow and the fifth framed",),# Guest post, 2026-10-07: Chunxiao Wang (Assay / nautilus-compass)# answered our practitioner invitation in #1118 + #1138. Every measurement@@ -172,6 +184,8 @@ PAGES: dict[str, dict[str, str]] = {title="Six eval scars from a production agent-memory system",description="Guest recipes from Chunxiao Wang (Assay): topic slicing beat token windows on recall@1 by double digits, a 0.87 classifier nearly matched by a 0.78 always-yes baseline, and the mask bug that scored every query zero. Plus the corpus gap we checked and confirmed.",published="2026-10-07", updated="2026-10-11",+ og_image="og/blog-eval-scars-from-agent-memory.png",+ og_image_alt="Cover: “Six eval scars from production agent memory” in display type beside the post’s square-matrix art, labelled with its figures. The art: six clusters of diagonal scratch marks made of small squares, five in ink and one in yellow",),# Exploit-hub A/B post, 2026-10-08. Every number traces to the# sealed-truth eval (/root/eval-holdout/EVAL-REPORT.md): 23744 route/@@ -183,6 +197,8 @@ PAGES: dict[str, dict[str, str]] = {title="The exploit hub, tested on three fresh CVEs",description="We ran the exploit corpus against three CVEs it had never seen: one query named the exact route and field for one, an in-corpus 2026 exploit named the exact denylist segments the vendor shipped for another, and the third proved the corpus adds nothing when its stratum is empty. Losses included.",published="2026-10-08", updated="2026-10-08",+ og_image="og/blog-the-exploit-hub.png",+ og_image_alt="Cover: “The exploit hub, tested on three fresh CVEs” in display type beside the post’s square-matrix art, labelled with its figures. The art: a network of small square nodes; a yellow hub linked to 36 yellow nodes, an ink node with one link, and an empty square frame with none",),"blog/index.html": dict(label="Blog", kind="collection",diff --git a/tools/packs/xerj-blogposts/posts/eval-scars-from-agent-memory.json b/tools/packs/xerj-blogposts/posts/eval-scars-from-agent-memory.jsonindex 73fee7f06..efd4aed0f 100644--- a/tools/packs/xerj-blogposts/posts/eval-scars-from-agent-memory.json+++ b/tools/packs/xerj-blogposts/posts/eval-scars-from-agent-memory.json@@ -4,5 +4,5 @@"description": "Guest recipes from Chunxiao Wang (Assay): topic slicing beat token windows on recall@1 by double digits, a 0.87 classifier nearly matched by a 0.78 always-yes baseline, and the mask bug that scored every query zero. Plus the corpus gap we checked and confirmed.","published": "2026-10-07","url": "https://xerj.org/blog/eval-scars-from-agent-memory",- "body": "← ALL POSTS · /BLOG \nGUEST CONTRIBUTION · AGENT MEMORY · RETRIEVAL EVAL · ISSUES #1118 + #1138 # SIX SCARS FROM PRODUCTION # AGENT-MEMORY EVAL, FROM SOMEONE BLEEDING THERE. \nGUEST · CHUNXIAO WANG, ASSAY / NAUTILUS-COMPASS · EDITED AND PUBLISHED BY XERJ ENGINEERING · 2026-10-07 \nWHY THIS POST EXISTS a guest answered our invitation with receipts We asked practitioners what a search engine built for agents is still missing. Chunxiao Wang of Assay ( nautilus-compass ) answered twice, in #1118 and #1138 , with six patterns from a production agent-memory system. Each one cost a real failure, which is the only currency eval advice should trade in. This post is their material, credited, with two additions from our side: where their patterns rhyme with how this repo already gates itself, and the honest answer to the question they asked back, namely where our corpus registry is thin against their workload. It is thin in exactly the way they guessed, and we checked before saying so. \nSix patterns, three themes: what changes when the corpus is the agent's own history, what an eval must freeze before anyone runs anything, and the arithmetic mistakes that make weak systems look strong. Every number in §01 through §04 is theirs. §05 is ours. \n## 01 · THE CORPUS IS THE AGENT'S OWN HISTORY \nRetrieving from a reference corpus and retrieving from session memory are different problems, and standard benchmarks cannot see the difference. Three of Chunxiao's observations carry it: \nThe observation \nTheir measurement \n\nthe query is a situation, not a question \nsession-memory queries arrive as imperatives shaped by task state (\"deploy failed again with 401\"), not information needs (\"how do 401s work?\"). Hybrid dense+sparse retrieval handled these better than dense-only in their runs, but the bigger effect was chunking \n\ntopic slicing beat token-window slicing \nchunking session memory by semantic topic instead of token window moved recall@1 by double digits in Chinese, and reversed the ranking of two embedding models relative to their MTEB scores. Benchmark-ledger performance did not predict in-distribution behavior on their session data \n\na failure hit is worth more than a manual hit \nwhen the corpus is the agent's own memory, a hit that surfaces that you already tried this and it failed beats a hit that surfaces the manual page. Nobody scores that today, in their words \n\nAnd the failure mode that hides from every dashboard: a fill_diagonal -style masking bug that excludes gold pairs from the candidate pool turns a retrieval eval into a constant-zero or constant-pass exercise. They shipped one. The test costs nothing: if every query scores zero hits, suspect your mask, not your model. \n## 02 · FOUR EVAL PATTERNS, EACH PAID FOR \nThe second contribution (#1118) is four patterns from building a verified-memory system. We keep them in their original order because the order is the story: distribution, leakage, calibration, variance. \nPattern \nTheir measurement \nThe recipe \n\nbenchmark scores do not transfer to your traffic \nbakeoff of bge-m3 vs Qwen3-embedding-0.6B on their own corpus (agent session memory, mixed zh/en): the MTEB-favored model gained +0.46pt recall@5 overall, under their transfer threshold, while the incumbent won recall@1 and the Chinese slices by +2.75pt \nre-rank embedders on a frozen sample of your own traffic before switching; public benchmarks are priors, not verdicts \n\ngroup by qid or your eval leaks \ncross-item contamination (train items from the same task family appearing in eval retrieval context) silently inflated their early numbers \nevery eval split is grouped by question or task id, never by row \n\nthree-valued retrieval beats binary \ntheir judge emits an unclear state with a confidence: Brier 0.069 with honest abstention, and the 2/27 abstentions in their latest public measurement were exactly the boundary rows \nlet the retriever say don't know; score abstention as part of calibration (ECE), not as failure \n\nto measure system drift, cache the inputs \nsame-day score swings of 0.93 → 0.30 → 0.90 looked like model drift; replaying cached inputs through the judge K times showed 0.0pp judge-side variance , all of it in the generated side \ntwo-stage variance attribution, replay-cached-input vs regenerate-then-judge, before blaming retrieval or judging \n\n## 03 · PREREGISTER, LABEL, BASELINE, PAIR \nThe third theme (#1138, recipe 2) is the discipline around the run itself, stated in four rules: \nThe rule \nThe scar behind it \n\nfreeze criteria before running: metric, threshold, stop-loss, and the exact verifiable artifact each claim points to. Criteria may only move stricter mid-stream; loosening needs a fresh preregistration \nevery rule here is a re-learned lesson from a run that was allowed to move its own goalposts \n\nlabel every claim measured, inferred, or unverifiable. Inferred claims carry an upgrade path \nthe most useful eval artifact they ever published was a self-report reading \"2 of 4 gates FAIL on recompute\" . The visible FAILs are what made the green credible afterwards \n\ncompute the majority-class baseline before claiming accuracy \na 0.87 binary classifier that a 0.78 always-yes baseline nearly matched, n=60, confidence intervals overlapping. Below n=100 the CI belongs next to every headline number, or the number is not a claim, it is a mood \n\npaired designs beat unpaired by an order of magnitude of information \nsame frames, two checkpoints, McNemar. Their unpaired reading overstated an effect by roughly 4× versus the paired one \n\nThe affinity is not an accident. This repo gates its own releases the same way: bars written before the work, FAILs kept visible in the notes, and a CI job that re-checks every claim in a release section against the tree it describes. The rc.80 gates post is the same philosophy with our numbers in it: three gates passed, three failed, one split, all published. When Chunxiao writes that visible FAILs are what made the green credible, that is the operating manual we already live by, arriving independently from a different domain. That convergence is worth more than either set of numbers alone. \n## 04 · THE QUESTION THEY ASKED BACK, ANSWERED HONESTLY \nTheir price for the recipes was not a star. It was this: an honest issue pointing at where the index is thin against these patterns, especially session-memory slicing and failure-turn retention. We checked the registry rather than answering from memory. The answer, posted in full in #1138 : \n0 agent-trajectory corpora at the hub, of 99 live 0 session or trajectory entries in the backlog, of 132 file / section our chunk unit, not the semantic topic their recall@1 needs Three thin spots, stated plainly. First, the corpus they described, real multi-session agent trajectories with timestamps and failure turns kept verbatim, does not exist at our hub: zero of the 99 live manifests and none of the other 131 backlog entries are session or trajectory data. Nobody using our reference corpora can score what they score. Second, our indexer chunks at file and section boundaries and keeps a locator back to the source span; for session logs that is the wrong shape, and we have no semantic-topic slicer to offer instead. Third, our memory features store curated facts an agent chose to keep, with BM25 recall by default; there is no native tried-and-failed marker and no retention of turns that went nowhere. \nWhat we did about it: the wanted corpus now has a named slot in the hub backlog ( agent-session-trajectories , status planned, demand-anchored to their issue), and if Assay can donate sanitized multi-session trajectories with failure turns intact, the contribution path is open by pull request . It would be a first-of-kind corpus at the hub. Their offer to recompute our benchmarks independently stands unaccepted for now, and it should not: an independent recompute of any number we publish is always welcome. \nWhat is claimed here, and by whom. Claimed by our guest, from their production system, with artifacts public in their repository: every measurement in §01 through §03, quoted from issues #1118 and #1138 with light editing and linked at each section. Claimed by us: the registry counts in §04 (98 manifests, backlog 132 rows after the wishlist entry, 99 live, zero matching a session, trajectory, chat or conversation query), verified 2026-10-07 by running the hub's own validator against the corpus-hub branch; the description of our chunking and memory features, which is a description of the code as it ships. Not claimed: that we have independently reproduced any of their numbers, that our engine today serves their workload well, or that the planned corpus exists in any form. It does not. That is the point of the backlog entry. If you have production scars of your own: the same invitation stands. Open an issue with what it cost you and what you measured, and if it survives our check against reality, it gets the same treatment, your name on it. The engine itself is one binary with a quickstart , and the agent-memory recipe is the closest thing we ship today to what §01 describes wanting. \nGuest material verbatim in issues #1118 and #1138 · their repository: nautilus-compass · the honest-gap comment: #1138 comment · backlog entry PR #1210 · related: the rc.80 gates · the Corpus Hub · the agent-memory recipe ."+ "body": "← ALL POSTS · /BLOG \nGUEST CONTRIBUTION · AGENT MEMORY · RETRIEVAL EVAL · ISSUES #1118 + #1138 # SIX SCARS FROM PRODUCTION # AGENT-MEMORY EVAL, FROM SOMEONE BLEEDING THERE. \nGUEST · CHUNXIAO WANG, ASSAY / NAUTILUS-COMPASS · EDITED AND PUBLISHED BY XERJ ENGINEERING · 2026-10-07 \nWHY THIS POST EXISTS a guest answered our invitation with receipts We asked practitioners what a search engine built for agents is still missing. Chunxiao Wang of Assay ( nautilus-compass ) answered twice, in #1118 and #1138 , with six patterns from a production agent-memory system. Each one cost a real failure, which is the only currency eval advice should trade in. This post is their material, credited, with two additions from our side: where their patterns rhyme with how this repo already gates itself, and the honest answer to the question they asked back, namely where our corpus registry is thin against their workload. It is thin in exactly the way they guessed, and we checked before saying so. \nSix patterns, three themes: what changes when the corpus is the agent's own history, what an eval must freeze before anyone runs anything, and the arithmetic mistakes that make weak systems look strong. Every number in §01 through §04 is theirs. §05 is ours. \n## 01 · THE CORPUS IS THE AGENT'S OWN HISTORY \nRetrieving from a reference corpus and retrieving from session memory are different problems, and standard benchmarks cannot see the difference. Three of Chunxiao's observations carry it: \nThe observation \nTheir measurement \n\nthe query is a situation, not a question \nsession-memory queries arrive as imperatives shaped by task state (\"deploy failed again with 401\"), not information needs (\"how do 401s work?\"). Hybrid dense+sparse retrieval handled these better than dense-only in their runs, but the bigger effect was chunking \n\ntopic slicing beat token-window slicing \nchunking session memory by semantic topic instead of token window moved recall@1 by double digits in Chinese, and reversed the ranking of two embedding models relative to their MTEB scores. Benchmark-ledger performance did not predict in-distribution behavior on their session data \n\na failure hit is worth more than a manual hit \nwhen the corpus is the agent's own memory, a hit that surfaces that you already tried this and it failed beats a hit that surfaces the manual page. Nobody scores that today, in their words \n\nAnd the failure mode that hides from every dashboard: a fill_diagonal -style masking bug that excludes gold pairs from the candidate pool turns a retrieval eval into a constant-zero or constant-pass exercise. They shipped one. The test costs nothing: if every query scores zero hits, suspect your mask, not your model. \n## 02 · FOUR EVAL PATTERNS, EACH PAID FOR \nThe second contribution (#1118) is four patterns from building a verified-memory system. We keep them in their original order because the order is the story: distribution, leakage, calibration, variance. \nPattern \nTheir measurement \nThe recipe \n\nbenchmark scores do not transfer to your traffic \nbakeoff of bge-m3 vs Qwen3-embedding-0.6B on their own corpus (agent session memory, mixed zh/en): the MTEB-favored model gained +0.46pt recall@5 overall, under their transfer threshold, while the incumbent won recall@1 and the Chinese slices by +2.75pt \nre-rank embedders on a frozen sample of your own traffic before switching; public benchmarks are priors, not verdicts \n\ngroup by qid or your eval leaks \ncross-item contamination (train items from the same task family appearing in eval retrieval context) silently inflated their early numbers \nevery eval split is grouped by question or task id, never by row \n\nthree-valued retrieval beats binary \ntheir judge emits an unclear state with a confidence: Brier 0.069 with honest abstention, and the 2/27 abstentions in their latest public measurement were exactly the boundary rows \nlet the retriever say don't know; score abstention as part of calibration (ECE), not as failure \n\nto measure system drift, cache the inputs \nsame-day score swings of 0.93 → 0.30 → 0.90 looked like model drift; replaying cached inputs through the judge K times showed 0.0pp judge-side variance , all of it in the generated side \ntwo-stage variance attribution, replay-cached-input vs regenerate-then-judge, before blaming retrieval or judging \n\n## 03 · PREREGISTER, LABEL, BASELINE, PAIR \nThe third theme (#1138, recipe 2) is the discipline around the run itself, stated in four rules: \nThe rule \nThe scar behind it \n\nfreeze criteria before running: metric, threshold, stop-loss, and the exact verifiable artifact each claim points to. Criteria may only move stricter mid-stream; loosening needs a fresh preregistration \nevery rule here is a re-learned lesson from a run that was allowed to move its own goalposts \n\nlabel every claim measured, inferred, or unverifiable. Inferred claims carry an upgrade path \nthe most useful eval artifact they ever published was a self-report reading \"2 of 4 gates FAIL on recompute\" . The visible FAILs are what made the green credible afterwards \n\ncompute the majority-class baseline before claiming accuracy \na 0.87 binary classifier that a 0.78 always-yes baseline nearly matched, n=60, confidence intervals overlapping. Below n=100 the CI belongs next to every headline number, or the number is not a claim, it is a mood \n\npaired designs beat unpaired by an order of magnitude of information \nsame frames, two checkpoints, McNemar. Their unpaired reading overstated an effect by roughly 4× versus the paired one \n\nThe affinity is not an accident. This repo gates its own releases the same way: bars written before the work, FAILs kept visible in the notes, and a CI job that re-checks every claim in a release section against the tree it describes. The rc.80 gates post is the same philosophy with our numbers in it: three gates passed, three failed, one split, all published. When Chunxiao writes that visible FAILs are what made the green credible, that is the operating manual we already live by, arriving independently from a different domain. That convergence is worth more than either set of numbers alone. \n## 04 · THE QUESTION THEY ASKED BACK, ANSWERED HONESTLY \nTheir price for the recipes was not a star. It was this: an honest issue pointing at where the index is thin against these patterns, especially session-memory slicing and failure-turn retention. We checked the registry rather than answering from memory. The answer, posted in full in #1138 : \n0 agent-trajectory corpora at the hub, of 99 live 0 session or trajectory entries in the backlog, of 132 file / section our chunk unit, not the semantic topic their recall@1 needs Three thin spots, stated plainly. First, the corpus they described, real multi-session agent trajectories with timestamps and failure turns kept verbatim, does not exist at our hub: zero of the 99 live manifests and none of the other 131 backlog entries are session or trajectory data. Nobody using our reference corpora can score what they score. Second, our indexer chunks at file and section boundaries and keeps a locator back to the source span; for session logs that is the wrong shape, and we have no semantic-topic slicer to offer instead. Third, our memory features store curated facts an agent chose to keep, with BM25 recall by default; there is no native tried-and-failed marker and no retention of turns that went nowhere. \nWhat we did about it: the wanted corpus now has a named slot in the hub backlog ( agent-session-trajectories , status planned, demand-anchored to their issue), and if Assay can donate sanitized multi-session trajectories with failure turns intact, the contribution path is open by pull request . It would be a first-of-kind corpus at the hub. Their offer to recompute our benchmarks independently stands unaccepted for now, and it should not: an independent recompute of any number we publish is always welcome. \nWhat is claimed here, and by whom. Claimed by our guest, from their production system, with artifacts public in their repository: every measurement in §01 through §03, quoted from issues #1118 and #1138 with light editing and linked at each section. Claimed by us: the registry counts in §04 (98 manifests, backlog 132 rows after the wishlist entry, 99 live, zero matching a session, trajectory, chat or conversation query), verified 2026-10-07 by running the hub's own validator against the corpus-hub branch; the description of our chunking and memory features, which is a description of the code as it ships. Not claimed: that we have independently reproduced any of their numbers, that our engine today serves their workload well, or — as of the 2026-10-07 registry check above — that the planned corpus exists in any form. It did not then. See the update below for what landed after. UPDATE · 2026-10-11 the thin spot named in §04 has since been seeded The gap this post was honest about on 2026-10-07 — zero session or trajectory corpora at the hub — is no longer zero. Chunxiao Wang donated the corpus they described, by pull request and under their own name, exactly as the invitation said it would land: PR #1255 (merged 2026-10-08) contributed 70 sanitized multi-session agent trajectories with failure turns intact , licensed CC-BY-4.0, through the documented contribution path . A maintainer then published the signed record pack, pack-agent-session-trajectories-2026-10-09 (2026-10-09), with the publish wiring in #1264 . It is the first session-trajectory corpus at the hub. The §04 counts above are left exactly as measured on 2026-10-07 and are not restated: the honest record is that the number was zero when we checked, and this note is what changed. What has not yet landed is the independent eval: Chunxiao's offer to recompute answers against the seeded corpus — pre-registered criteria, three-state verdicts, an UNVERIFIABLE wall — stands on their side and is not yet published. We make no claim on its result. \nIf you have production scars of your own: the same invitation stands. Open an issue with what it cost you and what you measured, and if it survives our check against reality, it gets the same treatment, your name on it. The engine itself is one binary with a quickstart , and the agent-memory recipe is the closest thing we ship today to what §01 describes wanting. \nGuest material verbatim in issues #1118 and #1138 · their repository: nautilus-compass · the honest-gap comment: #1138 comment · backlog entry PR #1210 · related: the rc.80 gates · the Corpus Hub · the agent-memory recipe ."}diff --git a/tools/packs/xerj-blogposts/posts/the-corpus-hub.json b/tools/packs/xerj-blogposts/posts/the-corpus-hub.jsonindex 6e290c201..a218aab98 100644--- a/tools/packs/xerj-blogposts/posts/the-corpus-hub.json+++ b/tools/packs/xerj-blogposts/posts/the-corpus-hub.json@@ -4,5 +4,5 @@"description": "A public registry of corpora for AI agents: reference code pinned to SHAs, record packs ed25519-signed, every source licence-reviewed by a human. The measured use case (2.7× fewer output tokens), the tie we published anyway, and how to contribute yours by PR.","published": "2026-10-01","url": "https://xerj.org/blog/the-corpus-hub",- "body": "← ALL POSTS · /BLOG \nCORPUS HUB · PINNED · LICENCED · SIGNED · BUILT 2026-10-01 # THE CORPUS HUB: PINNED, LICENCED, SIGNED, # AND CONTRIBUTED BY PULL REQUEST \nXERJ ENGINEERING · PUBLISHED 2026-10-01 \nEvery team that points an AI agent at a body of knowledge re-solves the same four problems in private: which exact version of the sources, under what licence, whether the copy you have is the copy I have, and who vouches for the bytes. We are opening the answers as a registry: the Corpus Hub , a branch on our GitHub where corpora arrive as pull requests — source pinned to SHAs, licences reviewed by a human per source, record packs checksummed and ed25519-signed. It ships with four reference corpora, one signed daily pack, a measured use case, and one result we did not want and are publishing anyway. \nUpdate, 2026-10-06: the registry has outgrown this post. The launch four are now 100 live corpora (119 backlog rows, every live one named in the xerj-code skill's domain-selection table since rc.83) across seven categories — including non-IT domains mined from real demand: eCFR titles 12/26/29 and UK legislation — and the registry has its own site: browse every corpus with its licence verdict, pin, size, use case and measured retrieval scores at hub.xerj.org . What follows is the post as written at launch. \n01 · THE PROBLEM ## Agent context is everyone's private plumbing \nAsk five teams what their coding agent retrieves from and you get five ad-hoc pipelines: a folder of cloned repos of unknown vintage, a wiki dump, a vector store someone seeded months ago. Each works, privately, and none of them can be shared, because a shared corpus needs four things a folder does not carry: \nProperty \nWhy it matters \nHow the Hub does it \n\npinned \ntwo people must retrieve the same bytes or results are incomparable \nevery repo recorded at the full SHA the build used; moving a pin is its own PR with a reason \n\nlicenced \n\"where did this text come from and may I act on it\" is not optional the day someone pastes \na human review block per source — adapt-freely / read-don't-paste / mixed — not a detector's guess \n\nchecksummed \na corpus that silently drifts is a corpus that lies \npacks carry per-file SHA-256; corpus add refuses the pack whole on mismatch \n\nsigned \nchecksums prove integrity; only a signature proves authorship \ndetached ed25519 over SHA256SUMS, verified with --verify-sig before a record is trusted \n\nThat last row was the strangest finding of our pack work : of the vulnerability databases people pip-install and OCI-pull daily, none ships a signature a consumer can verify. The Hub does, because a corpus an agent reasons from is supply chain, and supply chain runs on authorship, not just integrity. \n02 · WHAT IS IN IT ## Two lanes: reference corpora and record packs \nReference corpora are other people's source code, cloned at pinned commits — the material for reference coding , where an agent reads how three other engines solved the problem before writing a fourth. Record packs are structured data you curate — the first is 1,950 Rust vulnerability advisories identity-resolved across osv.dev and RustSec, rebuilt and signed daily. One command each: \n$ xerj corpus add --from https://raw.githubusercontent.com/xerj-org/xerj/corpus-hub/tools/xerj-code/hub/xerj-storage.json $ xerj corpus index xerj-storage $ xerj code xerj-storage \"flush epoch crash recovery\" \ncorpus \nlane \ncontents \nuse it for \n\nxerj-search \nreference \nlucene, tantivy, quickwit, meilisearch, sonic, elasticsearch \nFTS, BM25, merge policy, ES wire semantics \n\nxerj-vector \nreference \nqdrant, usearch, instant-distance, hnswlib \nHNSW construction, quantisation, filtered kNN \n\nxerj-storage \nreference \nsled, fjall, redb \nWAL, crash recovery, compaction \n\nxerj-columnar \nreference \nclickhouse \ncolumnar layout, codecs, vectorised scans \n\nrust-vulns \npack \n1,950 identity-resolved advisories, signed, daily \nvulnerability precedent, affected-function lookup \n\nThe grouping rule is the one design decision that matters: corpora are built by domain, not by language or popularity. A corpus that contains everything retrieves like a search engine with no query. The four reference corpora exist because an agent working on merge policy and an agent working on HNSW construction need different precedent — the same study that shows retrieval winning also shows the win gated by having the right precedent, not more of it. \n03 · THE USE CASE, MEASURED ## Reference coding: the same solve rate, 2.7× fewer output tokens \nThe Hub's reason to exist is the measured one. In a controlled study — 16 tasks across five languages, on libraries written for the study so no training set contains them, each carrying a runtime contract the compiler cannot warn about — an agent with XERJ reference corpora and the same agent with native file access both solved 16/16 . The difference was what it cost: \n\nnative (grep + read files) \nXERJ reference corpora \n\ntasks solved \n16/16 \n16/16 \n\noutput tokens (median) \n26,477 \n9,982 (2.7× fewer) \n\ncost \n$3.27 \n$1.58 (2.1× cheaper) \n\nRead the fine print, because it is the same fine print the next section turns into a whole experiment: on the study's two memorised public controls (tantivy; valkey + memcached), retrieval was neutral-to-harmful — the model reproduces even a 256-value quantisation table from memory. The collapse happened exactly where the agent could not have known the answer — a runtime rule like \"seal before you read\", a recovery order no compiler will tell you. Without retrieval, the same model solved 1 of 21 such tasks and burned $21.9 flailing; with native grep but no retrieval focus, one corpus had ~1.06M input tokens of file content pulled into context. \nThe pack lane answers a differently-shaped question. Reviewing code and wondering whether a pattern has precedent: \n$ xerj code rust-vulns \"smallvec insert_many buffer overflow\" \nreturns the identity-resolved advisory — one record carrying the RustSec prose, the GHSA filing, the CVE, the affected functions, the fixed versions — instead of three half-records in two databases that never talk to each other. \n04 · THE MEASUREMENT WE DID NOT WANT ## We A/B-tested the corpus on our own code. It tied. \nBefore opening a hub we ran the experiment that would justify it: two blinded arms auditing XERJ's own engine for security vulnerabilities — a plain agent, and the same agent with the rust-vulns corpus and XERJ retrieval. Same task, same turn budget, same pinned binary, two independent runs per arm, and a manual gate that re-proved every finding dynamically before counting it. Result, published in full : \n\nplain agent \ncorpus + XERJ \n\nfindings filed \n20 \n18 \n\nconfirmed exploitable (dynamic PoC) \n13 \n16 \n\nfalse positives \n1 — its only CRITICAL claim \n0 \n\ncost / wall clock \n$65.72 · 99.7 min \n$70.55 · 89.8 min \n\nfindings traceable to the corpus \n— \nzero \n\nThe corpus arm was marginally better on precision and it was never queried — one unused sweep call in one run, zero in the other. A tie, published as a tie, and then diagnosed against the pack itself: 1,963 records of advisory prose, only 14% containing any code, nothing that retrieves against a code-shaped question like \"does slicing at a byte offset inside a multi-byte character have precedent?\" — while half the pack was published after 2025, so memorisation was never the whole story. The failure was surface (nothing made retrieval the natural move), then shape (prose cannot match code), and only then size. \nThat diagnosis is now the Hub's roadmap, with pre-registered predictions : the next pack generation is code-shaped — version-pinned vulnerable functions extracted at the fix commit, before/after regression tests as trigger pairs, a bug-class taxonomy seeded from the 33 confirmed defect classes the two arms found — because we measured 9.8 extractable vulnerable functions per advisory and that is what a reviewer's question can actually retrieve against. If the re-test still ties, that gets published too. A registry whose numbers you can check is worth more than a demo. \n05 · CONTRIBUTE ## One pull request, two lanes, a licence table \nThe Hub is a branch: corpus-hub on xerj-org/xerj. Branch from it, open a PR against it, and a stdlib-only validator checks the format on every push — manifest fields, full 40-hex SHAs, licence review blocks, recipe shape, no binaries, no private keys. The full guide with templates lives in /docs/corpus-hub and in the branch's CONTRIBUTING.md . What a reviewer will actually hold you to: \nrule \nwhat it means \n\na nameable domain \nfinish the sentence \"an agent working on ___ would query this for ___\" or it is not ready \n\nun-memorised content \nretrieval's measured win is on niche, private, internal, post-cutoff material — not on famous snippets a model knows by heart \n\nhuman licence review \na review block per source with one of three verdicts: adapt-with-attribution (Apache/MIT/BSD), approach-only (AGPL/SSPL/Elastic/BUSL/GPL — read the design, never paste), or mixed \n\nno binaries in git \nrecipes and manifests only; built packs are signed Release assets a maintainer attaches after review \n\nThe licence row is where we are least flexible, because we have been burned by our own tooling: our licence detector once reported Elasticsearch as Apache-2.0 — the AGPL text contains the phrase «an \"Apache License 2.0\" compatible license» — and a detector's false \"permissive\" is how restricted code ends up pasted into an Apache-2.0 project. So the CI checks that a human review block exists ; only the reviewer's own opened licence file makes it true. \nWhat is next for the Hub itself: the code-shaped pack generation above, and the pre-indexed half — a corpus you mount instead of indexing, so a consumer's CPU is not the price of entry. Both tracked in the open. If you have a domain worth retrieving from — engine internals, protocol specs, incident postmortems, hardware errata — the branch is open."+ "body": "← ALL POSTS · /BLOG \nCORPUS HUB · PINNED · LICENCED · SIGNED · BUILT 2026-10-01 # THE CORPUS HUB: PINNED, LICENCED, SIGNED, # AND CONTRIBUTED BY PULL REQUEST \nXERJ ENGINEERING · PUBLISHED 2026-10-01 \nEvery team that points an AI agent at a body of knowledge re-solves the same four problems in private: which exact version of the sources, under what licence, whether the copy you have is the copy I have, and who vouches for the bytes. We are opening the answers as a registry: the Corpus Hub , a branch on our GitHub where corpora arrive as pull requests — source pinned to SHAs, licences reviewed by a human per source, record packs checksummed and ed25519-signed. It ships with four reference corpora, one signed daily pack, a measured use case, and one result we did not want and are publishing anyway. \nUpdate, 2026-10-06: the registry has outgrown this post. The launch four are now 100 live corpora (119 backlog rows, every live one named in the xerj-code skill's domain-selection table since rc.83) across seven categories — including non-IT domains mined from real demand: eCFR titles 12/26/29 and UK legislation — and the registry has its own site: browse every corpus with its licence verdict, pin, size, use case and measured retrieval scores at hub.xerj.org . What follows is the post as written at launch. \n01 · THE PROBLEM ## Agent context is everyone's private plumbing \nAsk five teams what their coding agent retrieves from and you get five ad-hoc pipelines: a folder of cloned repos of unknown vintage, a wiki dump, a vector store someone seeded months ago. Each works, privately, and none of them can be shared, because a shared corpus needs four things a folder does not carry: \nProperty \nWhy it matters \nHow the Hub does it \n\npinned \ntwo people must retrieve the same bytes or results are incomparable \nevery repo recorded at the full SHA the build used; moving a pin is its own PR with a reason \n\nlicenced \n\"where did this text come from and may I act on it\" is not optional the day someone pastes \na human review block per source — adapt-freely / read-don't-paste / mixed — not a detector's guess \n\nchecksummed \na corpus that silently drifts is a corpus that lies \npacks carry per-file SHA-256; corpus add refuses the pack whole on mismatch \n\nsigned \nchecksums prove integrity; only a signature proves authorship \ndetached ed25519 over SHA256SUMS, verified with --verify-sig before a record is trusted \n\nThat last row was the strangest finding of our pack work : of the vulnerability databases people pip-install and OCI-pull daily, none ships a signature a consumer can verify. The Hub does, because a corpus an agent reasons from is supply chain, and supply chain runs on authorship, not just integrity. \n02 · WHAT IS IN IT ## Two lanes: reference corpora and record packs \nReference corpora are other people's source code, cloned at pinned commits — the material for reference coding , where an agent reads how three other engines solved the problem before writing a fourth. Record packs are structured data you curate — the first is 1,950 Rust vulnerability advisories identity-resolved across osv.dev and RustSec, rebuilt and signed daily. From manifest to answer, with a local node running ( --from reads a local file, so the manifest is downloaded first): \n$ curl -fsSLO https://raw.githubusercontent.com/xerj-org/xerj/corpus-hub/tools/xerj-code/hub/xerj-storage.json $ xerj corpus add --from ./xerj-storage.json $ xerj corpus index xerj-storage $ xerj code xerj-storage \"flush epoch crash recovery\" \ncorpus \nlane \ncontents \nuse it for \n\nxerj-search \nreference \nlucene, tantivy, quickwit, meilisearch, sonic, elasticsearch \nFTS, BM25, merge policy, ES wire semantics \n\nxerj-vector \nreference \nqdrant, usearch, instant-distance, hnswlib \nHNSW construction, quantisation, filtered kNN \n\nxerj-storage \nreference \nsled, fjall, redb \nWAL, crash recovery, compaction \n\nxerj-columnar \nreference \nclickhouse \ncolumnar layout, codecs, vectorised scans \n\nrust-vulns \npack \n1,950 identity-resolved advisories, signed, daily \nvulnerability precedent, affected-function lookup \n\nThe grouping rule is the one design decision that matters: corpora are built by domain, not by language or popularity. A corpus that contains everything retrieves like a search engine with no query. The four reference corpora exist because an agent working on merge policy and an agent working on HNSW construction need different precedent — the same study that shows retrieval winning also shows the win gated by having the right precedent, not more of it. \n03 · THE USE CASE, MEASURED ## Reference coding: the same solve rate, 2.7× fewer output tokens \nThe Hub's reason to exist is the measured one. In a controlled study — 16 tasks across five languages, on libraries written for the study so no training set contains them, each carrying a runtime contract the compiler cannot warn about — an agent with XERJ reference corpora and the same agent with native file access both solved 16/16 . The difference was what it cost: \n\nnative (grep + read files) \nXERJ reference corpora \n\ntasks solved \n16/16 \n16/16 \n\noutput tokens (median) \n26,477 \n9,982 (2.7× fewer) \n\ncost \n$3.27 \n$1.58 (2.1× cheaper) \n\nRead the fine print, because it is the same fine print the next section turns into a whole experiment: on the study's two memorised public controls (tantivy; valkey + memcached), retrieval was neutral-to-harmful — the model reproduces even a 256-value quantisation table from memory. The collapse happened exactly where the agent could not have known the answer — a runtime rule like \"seal before you read\", a recovery order no compiler will tell you. Without retrieval, the same model solved 1 of 21 such tasks and burned $21.9 flailing; with native grep but no retrieval focus, one corpus had ~1.06M input tokens of file content pulled into context. \nThe pack lane answers a differently-shaped question. Reviewing code and wondering whether a pattern has precedent: \n$ xerj code rust-vulns \"smallvec insert_many buffer overflow\" \nreturns the identity-resolved advisory — one record carrying the RustSec prose, the GHSA filing, the CVE, the affected functions, the fixed versions — instead of three half-records in two databases that never talk to each other. \n04 · THE MEASUREMENT WE DID NOT WANT ## We A/B-tested the corpus on our own code. It tied. \nBefore opening a hub we ran the experiment that would justify it: two blinded arms auditing XERJ's own engine for security vulnerabilities — a plain agent, and the same agent with the rust-vulns corpus and XERJ retrieval. Same task, same turn budget, same pinned binary, two independent runs per arm, and a manual gate that re-proved every finding dynamically before counting it. Result, published in full : \n\nplain agent \ncorpus + XERJ \n\nfindings filed \n20 \n18 \n\nconfirmed exploitable (dynamic PoC) \n13 \n16 \n\nfalse positives \n1 — its only CRITICAL claim \n0 \n\ncost / wall clock \n$65.72 · 99.7 min \n$70.55 · 89.8 min \n\nfindings traceable to the corpus \n— \nzero \n\nThe corpus arm was marginally better on precision and it was never queried — one unused sweep call in one run, zero in the other. A tie, published as a tie, and then diagnosed against the pack itself: 1,963 records of advisory prose, only 14% containing any code, nothing that retrieves against a code-shaped question like \"does slicing at a byte offset inside a multi-byte character have precedent?\" — while half the pack was published after 2025, so memorisation was never the whole story. The failure was surface (nothing made retrieval the natural move), then shape (prose cannot match code), and only then size. \nThat diagnosis is now the Hub's roadmap, with pre-registered predictions : the next pack generation is code-shaped — version-pinned vulnerable functions extracted at the fix commit, before/after regression tests as trigger pairs, a bug-class taxonomy seeded from the 33 confirmed defect classes the two arms found — because we measured 9.8 extractable vulnerable functions per advisory and that is what a reviewer's question can actually retrieve against. If the re-test still ties, that gets published too. A registry whose numbers you can check is worth more than a demo. \n05 · CONTRIBUTE ## One pull request, two lanes, a licence table \nThe Hub is a branch: corpus-hub on xerj-org/xerj. Branch from it, open a PR against it, and a stdlib-only validator checks the format on every push — manifest fields, full 40-hex SHAs, licence review blocks, recipe shape, no binaries, no private keys. The full guide with templates lives in /docs/corpus-hub and in the branch's CONTRIBUTING.md . What a reviewer will actually hold you to: \nrule \nwhat it means \n\na nameable domain \nfinish the sentence \"an agent working on ___ would query this for ___\" or it is not ready \n\nun-memorised content \nretrieval's measured win is on niche, private, internal, post-cutoff material — not on famous snippets a model knows by heart \n\nhuman licence review \na review block per source with one of three verdicts: adapt-with-attribution (Apache/MIT/BSD), approach-only (AGPL/SSPL/Elastic/BUSL/GPL — read the design, never paste), or mixed \n\nno binaries in git \nrecipes and manifests only; built packs are signed Release assets a maintainer attaches after review \n\nThe licence row is where we are least flexible, because we have been burned by our own tooling: our licence detector once reported Elasticsearch as Apache-2.0 — the AGPL text contains the phrase «an \"Apache License 2.0\" compatible license» — and a detector's false \"permissive\" is how restricted code ends up pasted into an Apache-2.0 project. So the CI checks that a human review block exists ; only the reviewer's own opened licence file makes it true. \nWhat is next for the Hub itself: the code-shaped pack generation above, and the pre-indexed half — a corpus you mount instead of indexing, so a consumer's CPU is not the price of entry. Both tracked in the open. If you have a domain worth retrieving from — engine internals, protocol specs, incident postmortems, hardware errata — the branch is open."}diff --git a/tools/packs/xerj-blogposts/posts/the-rc80-gates.json b/tools/packs/xerj-blogposts/posts/the-rc80-gates.jsonindex 3b78e5ea7..d526b258c 100644--- a/tools/packs/xerj-blogposts/posts/the-rc80-gates.json+++ b/tools/packs/xerj-blogposts/posts/the-rc80-gates.json@@ -4,5 +4,5 @@"description": "Ten features, eight measured gates: /_ask macro-F1 0.9975, xerj_map 20/30 to 0/30, calibration ECE 0.0088 pass; the judge, the flywheel and the decide head fail their bars, numbers published.","published": "2026-09-30","url": "https://xerj.org/blog/the-rc80-gates",- "body": "← ALL POSTS · /BLOG \nRELEASE GATES · TEN FEATURES · EIGHT GATES · RUNS 2026-09-30 # THE RC.81 GATES. # MEASURED IN PUBLIC, LOSSES LEFT IN. \nAUTHOR · XERJ ENGINEERING TEAM · PUBLISHED · 2026-09-30 \nTHE VERDICT split, measured, with the losses left in The program behind rc.81 shipped ten features, and every item that could carry a number was filed with a measured gate that binds at ship time. This post is the packaging half of that rule, and the scoreboard is a split. Three gates passed outright : /_ask turned 230 prompt/gold pairs into validated DSL at macro-F1 0.9975 with zero invalid queries and wall-clock p50 1.129 ms against a 300 ms budget; xerj_map took unknown-field 400s on MCP-shaped DSL from 20/30 guessed to 0/30 map-informed — a 100% reduction against a ≥80% bar; and the calibration layer took FiQA rerank ECE from 0.3109 raw to 0.0088 held-out against a ≤0.10 bar. Three failed with the numbers : the local judge loses to its own first stage everywhere it was measured (SciFact 0.7045 → 0.5967 ), the decision flywheel answers 19.72% of later traffic at ≥0.8 confidence against an 80% bar, and the tier-2 decide head misses its accuracy bars by 3 to 59 points. One split : stemming gained +0.0160/+0.0179 BM25 nDCG@10 (bar +0.01) and still left 15 zero-hit queries against a bar of ≤10 — and no stemmer can fix the 15. The failures filed four issues, #1091–#1094 . No number below was tuned after its gate was read. \nOne window, ten features, eight gates, all measured on 2026-09-30 against bars that were written down before the work started. The wins and the losses sit in the same tables, every figure links to the run that produced it, and the four defects the gates exposed are filed in the open tracker with the measurements that found them. That is the whole post. \n0.9975 macro-F1, /_ask 230 prompt/gold pairs, 0 invalid DSL, byte-identical across 3 runs. p50 1.129 ms; the bar allowed 300. 0/30 unknown-field 400s, with xerj_map down from 20/30 guessed field names. 100% reduction; the bar asked for 80%. The 5 silent 0-hits became issue #1093. 0.0088 held-out ECE, calibrated from 0.3109 raw on 3,600 held-out FiQA pairs, query-level split. Temperature scaling managed 0.3163 and is published beside it. 4 issues the failures filed #1091 hybrid cost, #1092 analyzer echo, #1093 silent 0-hit, #1094 flywheel freeze. Each carries its measurement. 01 · THE PROGRAM · TEN FEATURES, EACH WITH A GATE THAT BINDS ## THE RULE WAS WRITTEN ## BEFORE THE WORK. \nThe ten-item program to the first official release ( discussion #1054 , filed as issues #1055 – #1064 ) was planned as five items for rc.79, four for rc.81, and the open model for v1.1.0. They landed together in one window — PRs #1066–#1090 since the rc.78 cut — so the gates were run as one suite, and the artifacts call it the rc.81 gate suite. This post keeps that name and states the plan honestly: two planned releases, one measured window. \nThe plan's own packaging rule, verbatim: a benchmarks/<name> directory with raw results plus a blog post with the losses left in. Three items also carried a loses-does-not-ship rule in their issue text — the local judge must beat the shipped hybrid arm beyond the run spread or it does not ship; the decide head below its bar ships behind a flag with the numbers published; the FiQA judge row must clear 0.30. This post is where those rules get honoured. \nThe program · feature (PR) · its gate \nVerdict \n\nxerj_map MCP FieldSpec tool ( #1074 ) · unknown-field 400s down ≥80% \nPASS — 100% \n\nPOST /_ask + xerj_plan ( #1076 ) · macro-F1 ≥0.9, 0 invalid DSL, p50 ≤300 ms \nPASS \n\ntier-2 local decide head via candle ( #1072 , #1073 ) · SMS ≥0.95 acc, ECE ≤0.05, AG News ≥0.85 \nFAIL (quality) / PASS (latency) \n\nmax_tokens on every MCP search tool ( #1067 ) · response-size budget knob \nno measured gate in this suite \n\nstemming analyzer for profiler-marked prose ( #1070 ) · BM25 ≥+0.01; zero-hit ≤10 \nSPLIT — +0.0160/+0.0179 PASS; 15 zero-hit FAIL \n\nself-judged hits judge stage ( #1077 ) · beat hybrid beyond spread, or do not ship \nFAIL (quality) / PASS (cost) \n\ndecision-cache flywheel ( #1075 ) · ≥80% of later traffic from history at ≥0.8, acc ≥0.97 \nFAIL (share) / PASS (accuracy) \n\nreal _watcher evaluate-or-501 + autoindex --label ( #1082 ) · wrong-and-confident ≤5% at p≥0.8 \ngate not yet measured \n\ncalibration layer, p_cal ( #1080 ) · held-out ECE ≤0.10 \nPASS — 0.0088 \n\nopen xerj-decide-v1 model ( #1084 ) · deterministic exports, published eval card \nreported — card carries the weak numbers \n\nTwo rows say \"not measured\" on purpose. max_tokens is a response-size budget with no quality claim attached, and the _watcher detection gate (wrong-and-confident rate at p≥0.8) has not been run — the watcher's mechanics are test-pinned, its detection-quality number is not, and this post does not invent one. \nOne binary, one day. Every number in this post was measured on 2026-09-30 on release builds of main @ a179e1ef3 , version string v1.0.0-rc.78 — the rc.79/rc.81 tags are cut after the gates, not before them. The two exceptions are stated where they occur (the tier-2 feature build, the paid provider arm). \nThe plan: discussion #1054 · tracker #1065 · issues #1055 – #1064 · window PRs #1066–#1090 · gate PRs #1079 , #1085 , #1086 , #1087 , #1088 , #1090 . \n02 · THE PASSES · THREE GATES, CLEAN ## THREE PASSED. ## WITH THEIR NUMBERS. \nPASS 1 ## POST /_ask ## + ## xerj_plan ## : prompt in, validated DSL out \n#1056 · PRs #1076, #1079 The gate: ≥200 (prompt, gold result-set) pairs over public tabular datasets, result-set F1 ≥0.9, zero invalid DSL, and a latency budget. Measured: 230 pairs (72 USGS earthquakes / 78 exoplanets / 80 gapminder), macro-F1 0.9975 , 0 invalid DSL — every returned plan passes parse_request before it is returned — and byte-identical output across 3 runs . Wall-clock p50 1.129 ms , p99 2.852 ms, n=690 over loopback HTTP. The bar allowed 300 ms; the planner clears it by 265×. \nThe one arm not run: the token gate — output tokens per solved structured-query task ≤50% of agent-written DSL at equal solve rate — is designed, not measured . It needs a live agent harness with real model calls, and the cost of that run was not spent this window. The design sits in benchmarks/ask-plan and the claim is not made here. \nPASS 2 ## xerj_map ## : the field map an agent was guessing at \n#1055 · PRs #1074, #1086 Thirty MCP-shaped DSL queries against a reference-coding corpus, issued twice: once guessing field names from the prompt alone, once informed by the map. Guessed: 20 of 30 came back as unknown-field 400s, and only 5 returned hits. Map-informed: 0 of 30 unknown-field 400s, 29 of 30 with hits. A 100% reduction against the ≥80% bar . \nThe discovery is worth more than the pass. Five of the guessed queries did not error at all — a wrong field name inside term , match , range or exists is a silent 0-hit HTTP 200 . The gate counted 400s; the silent rows are the worse failure, and they are now issue #1093 . \nPASS 3 ## Calibration: ## p_cal ## beside every ## p_raw \n#1063 · PRs #1080, #1087 The baseline is the number this blog already published: on FiQA, a hosted judge probability of 0.93 meant relevance 34% of the time — ECE 0.3109 . The gate: held-out ECE ≤0.10 after fitting. Measured on 19,440 raw (query, doc, p_raw, gold) pairs rebuilt from 648 judged FiQA queries — $0.2125 of provider calls, 5,059,586 input tokens, zero failed calls — with a query-level split: the fit never sees any pair from a held-out query, which is the deployment shape. Isotonic held-out ECE 0.0088 . Temperature scaling, the comparison arm, managed 0.3163 and fails the same bar; both arms' ECE is published by the node at /_decide/_calibration . \nThe gate also audited the fitter. Writing an independent naive PAVA in Python found a knot-index bug in the benchmark's mirror — fixed there, and the shipped Rust module was verified against the naive implementation to 8.7e-19 at all 98 pooled points. The defect was in the measuring stick, not the product; both are now pinned. \nask-plan: benchmarks/ask-plan (gate table with the measured rows) and results/2026-09-30-latency-gate · map gate: PR #1086 (artifacts in the PR) · calibration: results/2026-09-30-pairlevel-fiqa · pairs and shortlists committed beside them. \n03 · THE FAILURES · READINGS, NOT EXCUSES ## THREE FAILED, ONE SPLIT. ## THE NUMBERS STAY IN. \n## Stemming (#1059): a real win and a real miss, in the same gate \nPaired arms on one node, differing only in the create-time analyzer. BM25 nDCG@10: SciFact 0.6572 → 0.6732 (+0.0160) , NFCorpus 0.3016 → 0.3195 (+0.0179) — both clear the +0.01 bar. Zero-hit NFCorpus queries: 25 → 15, against a bar of ≤10 . Stemming repaired exactly the ten plural and inflection queries (bagels, leeks, pineapples, turnips…). The fifteen that remain — Fosamax, Zoloft, eggnog, halibut, Yale — are single words whose stem also occurs nowhere in the corpus . No stemmer can lexically bridge a word that is absent in every inflected form. That half of the gate was unreachable by the feature it gated; only the vector arm can fix those queries, and that is the honest reading: the gate was half right to exist, and the miss is a boundary, not a bug in the stemmer. \nTwo adjacent records: GET _settings does not echo the analysis block and _analyze shows the standard path even when the stemmer is declared and provably applied by search — filed as issue #1092 rather than tuned around. Raw per-query outputs are committed. \n## The tier-2 decide head (#1057): misses its bars, ships behind the flag anyway \nDecide head, xerj-decide-v1 weights · measured on the wire \nAccuracy \nECE \nBar \nVerdict \n\nSMS held-out, noul · 1,574 items \n0.9193 \n0.287 \nacc ≥0.95, ECE ≤0.05 \nFAIL / FAIL \n\nAG News test, 4-way choice · 7,600 items, never trained on it \n0.2599 \n0.060 \n≥0.85 \nFAIL (chance = 0.25) \n\nBanking77 test, 77-way choice · 3,080 items \n0.1185 \n0.0717 \nreported, not gated \nreported \n\nLatency, 30 questions, 8 pinned cores \np50 8.4 ms \n≤50 ms \nPASS \n\nThe reading is in the issue's own rule: below the bar it ships behind a flag with the numbers published — --decide-mode local --decide-model-dir , tier 2 off unless asked for. The artifact is 28.5 MiB of F32 weights, not quantized (candle 0.9 has no quantized safetensors VarBuilder ; the PR documents that as deliberate), well under the 300 MB download budget — and there is no download at all: the tier loads a local directory and makes no egress. Tier 1, the history vote, stays ahead of tier 2 on every home-ground row: Banking77 0.8289 vs 0.1185 , SMS 0.9848 vs 0.9193 , AG News 0.9182 vs 0.2599 . That is the design working: where labelled history exists, the vote wins; where it does not, tier 2 exists at all — and its zero-shot boundary is now a published number instead of a hope. \nThe gate's second catch is the more useful one. Measuring tier 2 on the stock release binary found that --decide-mode local was a no-op there: release binaries were built without the decide-local cargo feature, the node logged a warning, and the ladder quietly stayed on the history vote. Every tier-2 number above was therefore measured on a feature build of the same commit, with the no-op evidence committed ( stock-binary-noop.txt ). The fix is release engineering, not code: the shipped build must carry the feature — the runtime switch stays opt-in — and it is queued ahead of the cut. A gate that cannot reach the feature it is gating is measuring the build, and that is exactly what it caught. ## The flywheel (#1061): frozen as shipped, 4× under the bar even when forced \nTwo arms, both worse than the bar. As shipped : 2,000 fill requests fired, and only 12 answers were ever cached — write-back fires only when tier 2 answers, and after the first cached document the history vote returns BM25 support for ~99.4% of traffic, so tier 2 stops answering and the cache freezes at 12 documents. That is a defect, filed as issue #1094 , and it makes the gate's premise unreachable through the ladder itself. The premise forced — 2,000 gold answers cached in the write-back's own document shape — later traffic answered from history at ≥0.8 confidence: 19.72% against the 80% bar, four times under , at accuracy 0.9953 against the 0.97 bar, which met . The reading, not the excuse: the confident slice is easily accurate enough, but 2,000 answers at k=10 only clear 0.8 confidence on a fifth of Banking77-shaped traffic. The share bar needs a far larger cache, or the hybrid neighbours, before it is reachable — and until #1094 is fixed the cache cannot grow at all. \n## The local judge (#1060): loses to its own first stage, everywhere measured \nnDCG@10 · 3 runs each, spread 0.0000 \nFirst stage \n+ judge \nΔ \nBar \nVerdict \n\nSciFact · 300 q \n0.7045 hybrid \n0.5967 \n−0.1077 \nbeat hybrid beyond spread \nFAIL \n\nNFCorpus · 323 q \n0.3419 hybrid \n0.2924 \n−0.0495 \nbeat hybrid beyond spread \nFAIL \n\nFiQA · 648 q \n0.2382 BM25 \n0.1650 \n−0.0732 \n≥0.30 (hosted Jev: 0.3638) \nFAIL \n\nAdded latency, p50, top-30 \n+0.7 ms / +0.5 ms / +0.6 ms \n≤40 ms \nPASS \n\nWhy it loses, plainly: the shipped judge is a lexical scorer — window-local BM25 saturation with page-local IDF, no semantic signal. Both first stages already read those words: BM25 ranked by them, hybrid fused them with MiniLM vectors. Re-reading the same words over the page with page-local statistics destroys information the first stages had — corpus-level IDF and the vector arm — so the judged order is a worse order. The effect is biggest exactly where the first stage is strongest. The gate arms carried no min_p , because a threshold can only remove recall and cannot be part of a passing attempt; an informational arm with min_p: 0.5 kept 0.1 of 30 hits per page on SciFact and emptied 274 of 300 queries. The stage stays what PR #1077 shipped: opt-in, named lexical in every response, no quality claim — its own loses-does-not-ship rule, honoured. One cell could not be measured at all: hybrid+judge on FiQA, because the hybrid first stage runs 13–26 s per query on the 57,638-doc index — issue #1091 , found by this gate. \nStemming: results/2026-09-30-stemming-1059 (PR #1085) · decide + flywheel: PR #1088 — benchmarks/decisions-as-retrieval/results/2026-09-30-rc80gates/ , incl. run-2026-09-30.md and stock-binary-noop.txt · judge: PR #1090 — benchmarks/beir-hybrid/results/2026-09-30-judge-gate/ , incl. the aborted FiQA hybrid log behind #1091. \n04 · THE NEXT WINDOW · WHAT rc.81 DOES ABOUT EACH LOSS ## EVERY LOSS HAS ## A FILED ISSUE. \nA failed gate that files nothing is a buried gate. These filed four, plus two direction changes the numbers forced: \nThe loss \nThe filed work \n\nsilent 0-hit on unknown fields (found by the map gate) \n#1093 — QueryError::UnknownField is constructed nowhere today; wrong fields must 4xx, not 200-empty \n\nflywheel freeze (12 of 2,000 cached) \n#1094 — fix the write-back rule so the cache can grow; the 19.72% share measurement is then re-run on a live cache \n\nhybrid 13–26 s/query on FiQA (blocked the judge's FiQA hybrid cell) \n#1091 — first-stage cost on the 57,638-doc index; the hosted-comparison row needs it \n\nanalyzer invisible in settings/_analyze \n#1092 — echo the declared analysis block \n\nlexical judge loses everywhere \nthe model arm: a judge that ships must be the decide-local cross-encoder head, beating these same hybrid numbers under this same harness — the bar does not move \n\nthe 15 unstemmable zero-hit queries \nthe vector arm — the only repair for words absent from the corpus in every inflected form \n\nTwo more states, stated so nobody has to guess. The open model: xerj-decide-v1 is 7,419,907 parameters , its exports are byte-identical across runs, its eval card carries the weak numbers above in it, and the Hugging Face publication is staged behind an operator credential — no URL is claimed until the page exists . And the decide-head build fix: the shipped binary must carry the decide-local feature so the opt-in flag is not a no-op; that change rides ahead of the cut. \nWhat we claim, and what we do not. Claimed: every number on this page, verbatim from the gate runs linked beside it, on release builds of main @ a179e1ef3 measured 2026-09-30 — the tier-2 numbers on the feature build of that commit, the calibration pairs on one paid provider run whose determinism bound is stated in its provenance file. Not claimed: any quality for the judge stage, any quality for tier 2 beyond the published numbers, any token-saving result for xerj_plan (designed, not run), any detection-quality figure for _watcher (gate not yet measured), and any published URL for the model. The gates were written before the work; three said yes, three said no, one split; the repo carries all eight verdicts and the four issues the noes filed. If you want to check a number rather than read about it: each gate directory above carries its raw per-query outputs, its manifest and its harness, and the four gate PRs (#1086, #1088, #1090, and #1085's merged artifacts) are diffable line by line. The engine itself is one binary with a quickstart , the playground needs no install, and the benchmarks hub keeps the standing boards. The nearest prior chapter is the Jev verdict — the 0.3109 ECE this window's calibration gate finally answered. \nEverything on this page: the gate runs cited section by section above · program #1054 · gate PRs #1079 #1085 #1086 #1087 #1088 #1090 · issues #1091 – #1094 · related: the Jev verdict · the benchmarks hub ."+ "body": "← ALL POSTS · /BLOG \nRELEASE GATES · TEN FEATURES · EIGHT GATES · RUNS 2026-09-30 # THE RC.80 GATES. # MEASURED IN PUBLIC, LOSSES LEFT IN. \nAUTHOR · XERJ ENGINEERING TEAM · PUBLISHED · 2026-09-30 \nTHE VERDICT split, measured, with the losses left in The program behind rc.80 shipped ten features, and every item that could carry a number was filed with a measured gate that binds at ship time. This post is the packaging half of that rule, and the scoreboard is a split. Three gates passed outright : /_ask turned 230 prompt/gold pairs into validated DSL at macro-F1 0.9975 with zero invalid queries and wall-clock p50 1.129 ms against a 300 ms budget; xerj_map took unknown-field 400s on MCP-shaped DSL from 20/30 guessed to 0/30 map-informed — a 100% reduction against a ≥80% bar; and the calibration layer took FiQA rerank ECE from 0.3109 raw to 0.0088 held-out against a ≤0.10 bar. Three failed with the numbers : the local judge loses to its own first stage everywhere it was measured (SciFact 0.7045 → 0.5967 ), the decision flywheel answers 19.72% of later traffic at ≥0.8 confidence against an 80% bar, and the tier-2 decide head misses its accuracy bars by 3 to 59 points. One split : stemming gained +0.0160/+0.0179 BM25 nDCG@10 (bar +0.01) and still left 15 zero-hit queries against a bar of ≤10 — and no stemmer can fix the 15. The failures filed four issues, #1091–#1094 . No number below was tuned after its gate was read. \nOne window, ten features, eight gates, all measured on 2026-09-30 against bars that were written down before the work started. The wins and the losses sit in the same tables, every figure links to the run that produced it, and the four defects the gates exposed are filed in the open tracker with the measurements that found them. That is the whole post. \n0.9975 macro-F1, /_ask 230 prompt/gold pairs, 0 invalid DSL, byte-identical across 3 runs. p50 1.129 ms; the bar allowed 300. 0/30 unknown-field 400s, with xerj_map down from 20/30 guessed field names. 100% reduction; the bar asked for 80%. The 5 silent 0-hits became issue #1093. 0.0088 held-out ECE, calibrated from 0.3109 raw on 3,600 held-out FiQA pairs, query-level split. Temperature scaling managed 0.3163 and is published beside it. 4 issues the failures filed #1091 hybrid cost, #1092 analyzer echo, #1093 silent 0-hit, #1094 flywheel freeze. Each carries its measurement. 01 · THE PROGRAM · TEN FEATURES, EACH WITH A GATE THAT BINDS ## THE RULE WAS WRITTEN ## BEFORE THE WORK. \nThe ten-item program to the first official release ( discussion #1054 , filed as issues #1055 – #1064 ) was planned as five items for rc.79, four for rc.80, and the open model for v1.1.0. They landed together in one window — PRs #1066–#1090 since the rc.78 cut — so the gates were run as one suite, and the artifacts call it the rc.80 gate suite. This post keeps that name and states the plan honestly: two planned releases, one measured window. \nThe plan's own packaging rule, verbatim: a benchmarks/<name> directory with raw results plus a blog post with the losses left in. Three items also carried a loses-does-not-ship rule in their issue text — the local judge must beat the shipped hybrid arm beyond the run spread or it does not ship; the decide head below its bar ships behind a flag with the numbers published; the FiQA judge row must clear 0.30. This post is where those rules get honoured. \nThe program · feature (PR) · its gate \nVerdict \n\nxerj_map MCP FieldSpec tool ( #1074 ) · unknown-field 400s down ≥80% \nPASS — 100% \n\nPOST /_ask + xerj_plan ( #1076 ) · macro-F1 ≥0.9, 0 invalid DSL, p50 ≤300 ms \nPASS \n\ntier-2 local decide head via candle ( #1072 , #1073 ) · SMS ≥0.95 acc, ECE ≤0.05, AG News ≥0.85 \nFAIL (quality) / PASS (latency) \n\nmax_tokens on every MCP search tool ( #1067 ) · response-size budget knob \nno measured gate in this suite \n\nstemming analyzer for profiler-marked prose ( #1070 ) · BM25 ≥+0.01; zero-hit ≤10 \nSPLIT — +0.0160/+0.0179 PASS; 15 zero-hit FAIL \n\nself-judged hits judge stage ( #1077 ) · beat hybrid beyond spread, or do not ship \nFAIL (quality) / PASS (cost) \n\ndecision-cache flywheel ( #1075 ) · ≥80% of later traffic from history at ≥0.8, acc ≥0.97 \nFAIL (share) / PASS (accuracy) \n\nreal _watcher evaluate-or-501 + autoindex --label ( #1082 ) · wrong-and-confident ≤5% at p≥0.8 \ngate not yet measured \n\ncalibration layer, p_cal ( #1080 ) · held-out ECE ≤0.10 \nPASS — 0.0088 \n\nopen xerj-decide-v1 model ( #1084 ) · deterministic exports, published eval card \nreported — card carries the weak numbers \n\nTwo rows say \"not measured\" on purpose. max_tokens is a response-size budget with no quality claim attached, and the _watcher detection gate (wrong-and-confident rate at p≥0.8) has not been run — the watcher's mechanics are test-pinned, its detection-quality number is not, and this post does not invent one. \nOne binary, one day. Every number in this post was measured on 2026-09-30 on release builds of main @ a179e1ef3 , version string v1.0.0-rc.78 — the rc.79/rc.80 tags are cut after the gates, not before them. The two exceptions are stated where they occur (the tier-2 feature build, the paid provider arm). \nThe plan: discussion #1054 · tracker #1065 · issues #1055 – #1064 · window PRs #1066–#1090 · gate PRs #1079 , #1085 , #1086 , #1087 , #1088 , #1090 . \n02 · THE PASSES · THREE GATES, CLEAN ## THREE PASSED. ## WITH THEIR NUMBERS. \nPASS 1 ## POST /_ask ## + ## xerj_plan ## : prompt in, validated DSL out \n#1056 · PRs #1076, #1079 The gate: ≥200 (prompt, gold result-set) pairs over public tabular datasets, result-set F1 ≥0.9, zero invalid DSL, and a latency budget. Measured: 230 pairs (72 USGS earthquakes / 78 exoplanets / 80 gapminder), macro-F1 0.9975 , 0 invalid DSL — every returned plan passes parse_request before it is returned — and byte-identical output across 3 runs . Wall-clock p50 1.129 ms , p99 2.852 ms, n=690 over loopback HTTP. The bar allowed 300 ms; the planner clears it by 265×. \nThe one arm not run: the token gate — output tokens per solved structured-query task ≤50% of agent-written DSL at equal solve rate — is designed, not measured . It needs a live agent harness with real model calls, and the cost of that run was not spent this window. The design sits in benchmarks/ask-plan and the claim is not made here. \nPASS 2 ## xerj_map ## : the field map an agent was guessing at \n#1055 · PRs #1074, #1086 Thirty MCP-shaped DSL queries against a reference-coding corpus, issued twice: once guessing field names from the prompt alone, once informed by the map. Guessed: 20 of 30 came back as unknown-field 400s, and only 5 returned hits. Map-informed: 0 of 30 unknown-field 400s, 29 of 30 with hits. A 100% reduction against the ≥80% bar . \nThe discovery is worth more than the pass. Five of the guessed queries did not error at all — a wrong field name inside term , match , range or exists is a silent 0-hit HTTP 200 . The gate counted 400s; the silent rows are the worse failure, and they are now issue #1093 . \nPASS 3 ## Calibration: ## p_cal ## beside every ## p_raw \n#1063 · PRs #1080, #1087 The baseline is the number this blog already published: on FiQA, a hosted judge probability of 0.93 meant relevance 34% of the time — ECE 0.3109 . The gate: held-out ECE ≤0.10 after fitting. Measured on 19,440 raw (query, doc, p_raw, gold) pairs rebuilt from 648 judged FiQA queries — $0.2125 of provider calls, 5,059,586 input tokens, zero failed calls — with a query-level split: the fit never sees any pair from a held-out query, which is the deployment shape. Isotonic held-out ECE 0.0088 . Temperature scaling, the comparison arm, managed 0.3163 and fails the same bar; both arms' ECE is published by the node at /_decide/_calibration . \nThe gate also audited the fitter. Writing an independent naive PAVA in Python found a knot-index bug in the benchmark's mirror — fixed there, and the shipped Rust module was verified against the naive implementation to 8.7e-19 at all 98 pooled points. The defect was in the measuring stick, not the product; both are now pinned. \nask-plan: benchmarks/ask-plan (gate table with the measured rows) and results/2026-09-30-latency-gate · map gate: PR #1086 (artifacts in the PR) · calibration: results/2026-09-30-pairlevel-fiqa · pairs and shortlists committed beside them. \n03 · THE FAILURES · READINGS, NOT EXCUSES ## THREE FAILED, ONE SPLIT. ## THE NUMBERS STAY IN. \n## Stemming (#1059): a real win and a real miss, in the same gate \nPaired arms on one node, differing only in the create-time analyzer. BM25 nDCG@10: SciFact 0.6572 → 0.6732 (+0.0160) , NFCorpus 0.3016 → 0.3195 (+0.0179) — both clear the +0.01 bar. Zero-hit NFCorpus queries: 25 → 15, against a bar of ≤10 . Stemming repaired exactly the ten plural and inflection queries (bagels, leeks, pineapples, turnips…). The fifteen that remain — Fosamax, Zoloft, eggnog, halibut, Yale — are single words whose stem also occurs nowhere in the corpus . No stemmer can lexically bridge a word that is absent in every inflected form. That half of the gate was unreachable by the feature it gated; only the vector arm can fix those queries, and that is the honest reading: the gate was half right to exist, and the miss is a boundary, not a bug in the stemmer. \nTwo adjacent records: GET _settings does not echo the analysis block and _analyze shows the standard path even when the stemmer is declared and provably applied by search — filed as issue #1092 rather than tuned around. Raw per-query outputs are committed. \n## The tier-2 decide head (#1057): misses its bars, ships behind the flag anyway \nDecide head, xerj-decide-v1 weights · measured on the wire \nAccuracy \nECE \nBar \nVerdict \n\nSMS held-out, noul · 1,574 items \n0.9193 \n0.287 \nacc ≥0.95, ECE ≤0.05 \nFAIL / FAIL \n\nAG News test, 4-way choice · 7,600 items, never trained on it \n0.2599 \n0.060 \n≥0.85 \nFAIL (chance = 0.25) \n\nBanking77 test, 77-way choice · 3,080 items \n0.1185 \n0.0717 \nreported, not gated \nreported \n\nLatency, 30 questions, 8 pinned cores \np50 8.4 ms \n≤50 ms \nPASS \n\nThe reading is in the issue's own rule: below the bar it ships behind a flag with the numbers published — --decide-mode local --decide-model-dir , tier 2 off unless asked for. The artifact is 28.5 MiB of F32 weights, not quantized (candle 0.9 has no quantized safetensors VarBuilder ; the PR documents that as deliberate), well under the 300 MB download budget — and there is no download at all: the tier loads a local directory and makes no egress. Tier 1, the history vote, stays ahead of tier 2 on every home-ground row: Banking77 0.8289 vs 0.1185 , SMS 0.9848 vs 0.9193 , AG News 0.9182 vs 0.2599 . That is the design working: where labelled history exists, the vote wins; where it does not, tier 2 exists at all — and its zero-shot boundary is now a published number instead of a hope. \nThe gate's second catch is the more useful one. Measuring tier 2 on the stock release binary found that --decide-mode local was a no-op there: release binaries were built without the decide-local cargo feature, the node logged a warning, and the ladder quietly stayed on the history vote. Every tier-2 number above was therefore measured on a feature build of the same commit, with the no-op evidence committed ( stock-binary-noop.txt ). The fix is release engineering, not code: the shipped build must carry the feature — the runtime switch stays opt-in — and it is queued ahead of the cut. A gate that cannot reach the feature it is gating is measuring the build, and that is exactly what it caught. ## The flywheel (#1061): frozen as shipped, 4× under the bar even when forced \nTwo arms, both worse than the bar. As shipped : 2,000 fill requests fired, and only 12 answers were ever cached — write-back fires only when tier 2 answers, and after the first cached document the history vote returns BM25 support for ~99.4% of traffic, so tier 2 stops answering and the cache freezes at 12 documents. That is a defect, filed as issue #1094 , and it makes the gate's premise unreachable through the ladder itself. The premise forced — 2,000 gold answers cached in the write-back's own document shape — later traffic answered from history at ≥0.8 confidence: 19.72% against the 80% bar, four times under , at accuracy 0.9953 against the 0.97 bar, which met . The reading, not the excuse: the confident slice is easily accurate enough, but 2,000 answers at k=10 only clear 0.8 confidence on a fifth of Banking77-shaped traffic. The share bar needs a far larger cache, or the hybrid neighbours, before it is reachable — and until #1094 is fixed the cache cannot grow at all. \n## The local judge (#1060): loses to its own first stage, everywhere measured \nnDCG@10 · 3 runs each, spread 0.0000 \nFirst stage \n+ judge \nΔ \nBar \nVerdict \n\nSciFact · 300 q \n0.7045 hybrid \n0.5967 \n−0.1077 \nbeat hybrid beyond spread \nFAIL \n\nNFCorpus · 323 q \n0.3419 hybrid \n0.2924 \n−0.0495 \nbeat hybrid beyond spread \nFAIL \n\nFiQA · 648 q \n0.2382 BM25 \n0.1650 \n−0.0732 \n≥0.30 (hosted Jev: 0.3638) \nFAIL \n\nAdded latency, p50, top-30 \n+0.7 ms / +0.5 ms / +0.6 ms \n≤40 ms \nPASS \n\nWhy it loses, plainly: the shipped judge is a lexical scorer — window-local BM25 saturation with page-local IDF, no semantic signal. Both first stages already read those words: BM25 ranked by them, hybrid fused them with MiniLM vectors. Re-reading the same words over the page with page-local statistics destroys information the first stages had — corpus-level IDF and the vector arm — so the judged order is a worse order. The effect is biggest exactly where the first stage is strongest. The gate arms carried no min_p , because a threshold can only remove recall and cannot be part of a passing attempt; an informational arm with min_p: 0.5 kept 0.1 of 30 hits per page on SciFact and emptied 274 of 300 queries. The stage stays what PR #1077 shipped: opt-in, named lexical in every response, no quality claim — its own loses-does-not-ship rule, honoured. One cell could not be measured at all: hybrid+judge on FiQA, because the hybrid first stage runs 13–26 s per query on the 57,638-doc index — issue #1091 , found by this gate. \nStemming: results/2026-09-30-stemming-1059 (PR #1085) · decide + flywheel: PR #1088 — benchmarks/decisions-as-retrieval/results/2026-09-30-rc80gates/ , incl. run-2026-09-30.md and stock-binary-noop.txt · judge: PR #1090 — benchmarks/beir-hybrid/results/2026-09-30-judge-gate/ , incl. the aborted FiQA hybrid log behind #1091. \n04 · THE NEXT WINDOW · WHAT rc.81 DOES ABOUT EACH LOSS ## EVERY LOSS HAS ## A FILED ISSUE. \nA failed gate that files nothing is a buried gate. These filed four, plus two direction changes the numbers forced: \nThe loss \nThe filed work \n\nsilent 0-hit on unknown fields (found by the map gate) \n#1093 — QueryError::UnknownField is constructed nowhere today; wrong fields must 4xx, not 200-empty \n\nflywheel freeze (12 of 2,000 cached) \n#1094 — fix the write-back rule so the cache can grow; the 19.72% share measurement is then re-run on a live cache \n\nhybrid 13–26 s/query on FiQA (blocked the judge's FiQA hybrid cell) \n#1091 — first-stage cost on the 57,638-doc index; the hosted-comparison row needs it \n\nanalyzer invisible in settings/_analyze \n#1092 — echo the declared analysis block \n\nlexical judge loses everywhere \nthe model arm: a judge that ships must be the decide-local cross-encoder head, beating these same hybrid numbers under this same harness — the bar does not move \n\nthe 15 unstemmable zero-hit queries \nthe vector arm — the only repair for words absent from the corpus in every inflected form \n\nTwo more states, stated so nobody has to guess. The open model: xerj-decide-v1 is 7,419,907 parameters , its exports are byte-identical across runs, its eval card carries the weak numbers above in it, and the Hugging Face publication is staged behind an operator credential — no URL is claimed until the page exists . And the decide-head build fix: the shipped binary must carry the decide-local feature so the opt-in flag is not a no-op; that change rides ahead of the cut. \nWhat we claim, and what we do not. Claimed: every number on this page, verbatim from the gate runs linked beside it, on release builds of main @ a179e1ef3 measured 2026-09-30 — the tier-2 numbers on the feature build of that commit, the calibration pairs on one paid provider run whose determinism bound is stated in its provenance file. Not claimed: any quality for the judge stage, any quality for tier 2 beyond the published numbers, any token-saving result for xerj_plan (designed, not run), any detection-quality figure for _watcher (gate not yet measured), and any published URL for the model. The gates were written before the work; three said yes, three said no, one split; the repo carries all eight verdicts and the four issues the noes filed. If you want to check a number rather than read about it: each gate directory above carries its raw per-query outputs, its manifest and its harness, and the four gate PRs (#1086, #1088, #1090, and #1085's merged artifacts) are diffable line by line. The engine itself is one binary with a quickstart , the playground needs no install, and the benchmarks hub keeps the standing boards. The nearest prior chapter is the Jev verdict — the 0.3109 ECE this window's calibration gate finally answered. \nEverything on this page: the gate runs cited section by section above · program #1054 · gate PRs #1079 #1085 #1086 #1087 #1088 #1090 · issues #1091 – #1094 · related: the Jev verdict · the benchmarks hub ."}