Skip to content

docs(benchmarks): qualify typed-reply source evidence [sc-2866] #604

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
norvalbv merged 1 commit into from
Sep 6, 2026
Merged

docs(benchmarks): qualify typed-reply source evidence [sc-2866] #604

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions docs/benchmarks/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -58,6 +58,7 @@ Hand-written records that sit beside the ledger; each number in them names its p
- [`reviewer-yield-vs-diff-size.md`](reviewer-yield-vs-diff-size.md) — the review gate returns ~1 issue per run above ~1k LOC and the corpus never measures that regime (2026-08-22).
- [`corpus-growth.md`](corpus-growth.md) — how corpus rows are mined from telemetry and adapted.
- [`source-family-census.md`](source-family-census.md) — source-qualified candidate families and exact native initial-evidence coverage (zero reviewer calls).
- [`source-family-qualification-2026-09-06.md`](source-family-qualification-2026-09-06.md) — source-controlled reply extraction and unresolved delivery-boundary qualification (sc-2866).
- [`correctness-rebaseline-2026-09-04.md`](correctness-rebaseline-2026-09-04.md) — correctness re-baselined on the shipped gpt-5.6-sol pin: recall 0.92, clean-pass 0.60, and the precision loss sits entirely in the fix-pair decoys (sc-2494).
- [`experiments/`](experiments/) — one folder per investigation with its inputs, arms, and reports (latest: [`2026-08-24-codex-judge-probe`](experiments/2026-08-24-codex-judge-probe/README.md) — OpenAI codex-CLI models as gate judges on the round-1 corpus: gpt-5.6-sol@chunk:400 13/23 and gpt-5.6-terra@xhigh WHOLE-diff 11/23 vs sonnet's best 8/23, ~perfect checklist compliance across 840 tasks (839 terminal); chunk size and reasoning effort proved ALTERNATIVE strategies, and terra's effort curve is a cliff at xhigh. Prior: [`2026-08-23-scale-probe`](experiments/2026-08-23-scale-probe/README.md) — whole-diff vs chunked correctness review on 8 real 1k–7.2k-LOC diffs: chunking pooled ≥ whole-diff under all five mining/scoring rulesets tried (shipped ruleset: chunk:1000 8/23 vs whole 5/23 on decontaminated labels), but the ratio moves 1.375×–2.0× with the ruleset so no registered bar is treated as cleared — decision moved to the confirmation round; ~half the round's original labels were test-retest contamination, since fixed; haiku judges failed their arms; full correction record inside. Prior: [`2026-08-22-ship-attempts-research`](experiments/2026-08-22-ship-attempts-research/README.md)).

Expand Down
Loading
Loading