ruvnet.
Python
docs(ruvnet): record 2026-09-15 evidence
rUv390ef0640ede
diff --git a/docs/self-improvement/2026-09-15-evidence.json b/docs/self-improvement/2026-09-15-evidence.jsonnew file mode 100644index 0000000..548d117--- /dev/null+++ b/docs/self-improvement/2026-09-15-evidence.json@@ -0,0 +1,71 @@+{+ "schemaVersion": 1,+ "date": "2026-09-15",+ "decision": "accept",+ "baseline": "d38753c88aaab3fd958b2dc9802f3b858fb94cb0",+ "candidateCommit": "cc91910351c34789db941e29960b905f1e0db2f1",+ "candidate": "exact-change-feed-filters",+ "version": "0.3.0",+ "upstream": {+ "observedAt": "2026-09-15T10:24:49Z",+ "sources": [+ {"project":"ruflo","revision":"2602b642d92234c710ffbe96bfb33007d481ceab","license":"MIT","changedSincePrior":true,"incorporated":"Ruflo 3.42.0 release provenance plus ten constituent security, memory, learning, swarm, performance and observability records"},+ {"project":"metaharness","revision":"d5833dc6512ac1adeeef91a331c29055cd8a4dbb","license":"MIT","changedSincePrior":false},+ {"project":"ruvector","revision":"edaffffb3b85768eb1f3ec1f683b7f46f0506af4","license":"MIT","changedSincePrior":false},+ {"project":"agentdb","revision":"6cdefe451f6afe7395bc88d3587a0225d161089b","license":"MIT","changedSincePrior":false},+ {"project":"dream-machine","revision":"3edd426f6c9c4b1e80235f7447dc863e749345cc","license":"MIT","changedSincePrior":false,"incorporated":false},+ {"project":"autogenous","revision":"7bf327a9754ce798364dbee8b2825af42a421fd4","license":"MIT","changedSincePrior":false,"incorporated":false},+ {"project":"rvm","revision":"0973d77e5a9e99065376e8ef772a4a1d16e88c75","license":"MIT","changedSincePrior":false,"incorporated":false}+ ],+ "rufloCommitsReviewed": 18,+ "newRecordsIncorporated": 11,+ "xRuvIo":{"version":"0.7.2","routes":["/mcp","/chatgpt/mcp"],"changedSincePrior":false}+ },+ "heldOut":{"baseline":{"passed":0,"total":6},"candidate":{"passed":6,"total":6}},+ "tests":{"groupsPassed":10,"groupsFailed":0},+ "validation":{"plugin":true,"skillsPassed":4,"skillsFailed":0,"marketplaceSchemas":2,"packageFiles":18,"compressedBytes":19889,"unpackedBytes":59975,"lockfileSha256":"e7249650c8003ffec523bb8920bdae64eaeede844ae2a31acf43d85e18371ae3"},+ "benchmarks": {+ "raw":"2026-09-15-benchmarks.json",+ "pairs":15,+ "baseline":{"discoveryP95MedianMs":0.0013319999999978904,"coldCliP95MedianMs":71.5330570000001,"rssMedianBytes":47640576},+ "candidate":{"discoveryP95MedianMs":0.001221999999998502,"coldCliP95MedianMs":71.210771,"rssMedianBytes":47747072},+ "delta":{"discoveryP95Percent":-8.25825825822542,"coldCliP95Percent":-0.45054134901588844,"coldCliP95Ms":-0.3222860000001049,"rssBytes":106496},+ "interpretation":"All frozen regression gates passed. Timing deltas are not claimed as product speed improvements."+ },+ "metaharness":{"package":"[email protected]","scope":"plugins/ruvnet","hardConstraints":"6/6","baseline":{"harnessFit":60,"compileConfidence":65,"taskCoverage":84,"toolSafety":90,"memoryUsefulness":38},"candidate":{"harnessFit":60,"compileConfidence":65,"taskCoverage":84,"toolSafety":90,"memoryUsefulness":38}},+ "security": {+ "rufloPackage":"@claude-flow/[email protected]",+ "deepCodeScan":{"critical":0,"high":0,"medium":0,"low":0},+ "deepDependencyScan":{"critical":0,"high":0,"medium":0,"low":0},+ "npmAudit":{"critical":0,"high":0,"moderate":0,"low":0,"totalDependenciesReported":94},+ "secretScan":{"findings":0,"filesReportedByScanner":5},+ "manualTrackedFileCredentialPatternHits":0,+ "threatScan":{"indicators":0,"filesReportedByScanner":4},+ "trustBoundaries":["CLI positional arguments","MCP tool arguments","reviewed upstream JSON","host interpretation of annotations and linked evidence"],+ "residualRisks":["MCP annotations are advisory to hosts","the reviewed snapshot can become stale","upstream performance and evaluation claims were not independently reproduced","secret and STRIDE scanners reported narrow coverage"],+ "limitations":"Zero reported findings is not proof of complete security."+ },+ "memoryExperiment": {+ "catalogProjects":10,+ "changeRecords":17,+ "eligiblePublicRecords":27,+ "ruvectorRevision":"edaffffb3b85768eb1f3ec1f683b7f46f0506af4",+ "supportedApisInspected":["ruvector VectorDB with durable storage and metadata filters","@ruvector/wasm VectorDB plus metadata-sidecar adapter","ruvector-gnn-rerank CandidateReranker"],+ "ruvectorWasm":"deferred",+ "gnn":"rejected for this workload",+ "persistentLearning":"not implemented",+ "reason":"The sanitized corpus remains below the frozen 100-record minimum. At the pinned revision, WASM HNSW is inactive and raw WASM metadata does not round-trip; the adapter supplies a sidecar. Upstream GNN evidence is a clustered synthetic reranking result, while exact L2 is the documented quality ceiling and materially faster on that benchmark.",+ "futureGate":{"minimumRecords":100,"minimumTaskSuccessGainPercentagePoints":10,"p95MaximumMs":20,"memoryMaximumBytes":134217728,"requiredAblations":["exact filters","vector only","exact plus vector fusion","GNN rerank","shuffled-graph control"]}+ },+ "forecast": {+ "prediction":"Agent change ledgers and capability catalogs will increasingly expose server-side structured filters or stable pagination so hosts can retrieve provenance with deterministic cache keys.",+ "status":"prediction, not established result",+ "confidencePercent":80,+ "horizonDate":"2026-11-14",+ "criterion":"At least two primary agent-memory, MCP, or tool-registry implementations newly expose public structured filtering or stable pagination/continuation interfaces for provenance records.",+ "sources":["https://arxiv.org/abs/2604.15484","https://blog.modelcontextprotocol.io/posts/2026-07-28/","https://github.com/ruvnet/ruflo/releases/tag/v3.42.0"],+ "priorForecastsExpired":[]+ },+ "ci":{"run":"https://github.com/ruvnet/ruvnet/actions/runs/34959021726","status":"success","node22":"success","node24":"success"},+ "dashboard":"https://ruvnet-harness-evidence.ruv.chatgpt.site"+}diff --git a/docs/self-improvement/2026-09-15.md b/docs/self-improvement/2026-09-15.mdnew file mode 100644index 0000000..d94af01--- /dev/null+++ b/docs/self-improvement/2026-09-15.md@@ -0,0 +1,31 @@+# RuVnet harness improvement — 2026-09-15++## Decision++Accepted exact project and change-kind filters and published RuVnet entry point 0.3.0. `ruvnet changes`, `ruvnet_changes`, and the portable start skill can now narrow reviewed provenance without semantic inference. Filters are allow-listed from the reviewed snapshot, intersected, counted before limiting, and rejected on unknown or extra values. The companion remains offline, read only, and incapable of refreshing or learning by itself.++## Upstream review++Ruflo advanced 18 commits to [3.42.0](https://github.com/ruvnet/ruflo/releases/tag/v3.42.0) at `2602b642d92234c710ffbe96bfb33007d481ceab`. The snapshot incorporates the release plus ten constituent records covering weighted consensus, HNSW product-quantization dispatch, weighted RRF, near-duplicate memory, confidence separation, reward-aware consolidation, degraded fallback disclosure, lazy MMR, MCP governance and hive-mind capability authentication. The release says these changes have regression evidence; that upstream claim was not independently reproduced in this repository. MetaHarness, RuVector, AgentDB, Dream Machine, Autogenous and RVM did not advance from the previously recorded heads. All reviewed repositories declare MIT licenses. x.ruv.io still reports gateway 0.7.2 with `/chatgpt/mcp` and `/mcp`.++## MetaHarness loop and measurements++The frozen hypothesis was that exact structured filters would improve relevant provenance retrieval without expanding authority. Baseline `d38753c88aaab3fd958b2dc9802f3b858fb94cb0` passed 0/6 held-out checks; the candidate passed 6/6. Ten test groups, both marketplace schemas, all four portable skills, the MCP SDK subprocess/in-memory paths, package closure, and negative boundaries passed. MetaHarness 0.4.16 remained at 6/6 hard constraints; its readiness score did not change and is not used as proof of improvement.++Fifteen alternating baseline/candidate pairs used 10,000 discovery calls and 20 cold CLI starts per run. Median discovery p95 moved from 0.001332 ms to 0.001222 ms, cold CLI p95 from 71.533 ms to 71.211 ms, and RSS from 47,640,576 to 47,747,072 bytes. All frozen regression gates passed; sub-microsecond timing differences are noise, not a speed claim. Raw results are in [2026-09-15-benchmarks.json](2026-09-15-benchmarks.json).++The pinned Ruflo 3.25.6 deep code and dependency scans, npm audit, and manual tracked-file credential pattern check reported zero known findings. The secret scanner reported only five files and STRIDE only four, so coverage is limited and this is not a claim of complete security. Residual risks are stale reviewed data, upstream claims that were not locally reproduced, host treatment of advisory MCP annotations, and positional CLI ambiguity.++[Candidate commit](https://github.com/ruvnet/ruvnet/commit/cc91910351c34789db941e29960b905f1e0db2f1) passed [Node 22 and 24 CI](https://github.com/ruvnet/ruvnet/actions/runs/34959021726).++## RuVector, WASM and GNN decision++At pinned RuVector `edaffffb3b85768eb1f3ec1f683b7f46f0506af4`, supported surfaces include durable `VectorDB` search with metadata filters, `@ruvector/wasm` with an adapter-side metadata store, and `ruvector-gnn-rerank`. The same documentation states that WASM HNSW is not active, raw WASM metadata does not round-trip, and GNN reranking is not automatically wired into `VectorDB::search`. Its clustered synthetic GNN benchmark reports +10.4 percentage points over a noisy first stage, but exact L2 is the quality ceiling and much faster in that benchmark.++This run has 10 catalog projects plus 17 reviewed changes: 27 eligible public records, below the frozen 100-record minimum. Persistent vector memory, WASM acceleration, GNN training and persistent learning were therefore not implemented or claimed. The next eligible experiment must compare exact filters, vector-only retrieval, fused retrieval, GNN reranking and a shuffled-graph control on held-out tasks.++## Forecast++Prediction, 80% confidence by 2026-11-14: agent change ledgers and capability catalogs will increasingly expose server-side structured filters or stable pagination so hosts can retrieve provenance with deterministic cache keys. It passes only if at least two primary agent-memory, MCP or tool-registry implementations newly expose such public interfaces. Motivation comes from [vstash's measured hybrid retrieval and adaptive fusion](https://arxiv.org/abs/2604.15484), the [2026-07-28 MCP specification](https://blog.modelcontextprotocol.io/posts/2026-07-28/), and Ruflo's rapidly growing evaluated change set. No previous forecast horizon expired today.++Next hypothesis: add deterministic stable pagination only after the reviewed feed exceeds 20 records and a frozen held-out host task demonstrates that it prevents missed evidence without increasing stale-cache errors.