Official Codebase of 3DCodeVerse
Python
docs: the link rewrites that the retire commit's aborted git-add left behind
gaoypeng9b393e570338
diff --git a/harness/bench/paired_compare.py b/harness/bench/paired_compare.pyindex 2b0f47ca..c6693886 100644--- a/harness/bench/paired_compare.py+++ b/harness/bench/paired_compare.py@@ -1,7 +1,7 @@"""Paired analysis of a ``compare_backends`` battery: harness arm − one-shot arm per prompt.``bench/_compare_report.py`` tabulates arms; this answers the paper's question — *is the-harness lift separated from noise?* — the way ``docs/report.html`` §6 states it: every+harness lift separated from noise?* — the way the retired compare report framed it: everycomparison carries its paired standard error, a 95 % confidence interval and the exacttwo-sided sign test, and a comparison whose interval crosses zero is labelled``unsupported`` however good the mean looks.diff --git a/harness/docs/COST.md b/harness/docs/COST.mdindex 47d0aa2b..6b193406 100644--- a/harness/docs/COST.md+++ b/harness/docs/COST.md@@ -1226,7 +1226,7 @@ guard is now `len(failed_keys) < len(pool)`.-**Follow-up, same day, from the time audit (51 storm-day runs vs 52 baseline; scripts retired 2026-08-28 — findings preserved in `docs/TIME_AUDIT_2026-08-26.md`).**+**Follow-up, same day, from the time audit (51 storm-day runs vs 52 baseline; scripts retired 2026-08-28 — findings summarised in §28 below).**Three accelerations, all additive and on by default:*A caller clips the retry budget to what it can afford* (`ChatRequest.max_wait_s`, None = the@@ -1269,7 +1269,7 @@ in `raw` and `attempts` on the raised `ModelError`; `cost/instrument.py` copiesdefaulted so old rows load); `3dcv cost` / `bench/cost_report.py` add a per-key table and a`tries/call` column (`CostBucket.attempts_per_call`) whenever the ledger carries them.-## 28. Where the time goes — the 2026-08-26 audit (`docs/TIME_AUDIT_2026-08-26.md`)+## 28. Where the time goes — the 2026-08-26 audit51 storm-day runs against 52 baseline runs, every stage and every model call, scripts in`bench/time_audit/` (read-only over `bench/out`). The numbers that decide what to build next:diff --git a/harness/docs/EVAL.md b/harness/docs/EVAL.mdindex 2bba95cb..1fd955c4 100644--- a/harness/docs/EVAL.md+++ b/harness/docs/EVAL.md@@ -212,7 +212,7 @@ same contact sheets ("would a curator screenshot it / does it look like the thinThree aurora versions nobody would take for an aurora scored 0.78 / 0.94 / 0.94 with empty issuelists; opaque pastel discs for bokeh 0.92; a lifted purple wash for a nebula 0.82; a crispukiyo-e wave 0.59 under planner must items. The rubric scored the nouns of the brief being-present. `docs/GRAPHICS_LOOP.md` is the loop that fixes this (rubric `shader_v2`: likeness,+present. The graphics loop fixes this (rubric `shader_v2`: likeness,tonal range, an artefact checklist; reference photos via `bench/refs/<id>/`; `LikenessJudge`)and the ledger of turns; `bench/judge_calib_graphics.py` re-judges the corpus under two rubricsagainst the eye file and is the gate for switching the track default. The rule from §6 holds: