Skip to content

bench(reviewer-correctness): first checkpoint on the mined 90-row corpus - #297

Merged
norvalbv merged 1 commit into
mainfrom
bench/publish-correctness-90
Aug 1, 2026
Merged

bench(reviewer-correctness): first checkpoint on the mined 90-row corpus#297
norvalbv merged 1 commit into
mainfrom
bench/publish-correctness-90

Conversation

@norvalbv

@norvalbv norvalbv commented Aug 1, 2026

Copy link
Copy Markdown
Owner

PR E — stacked on #296. Publishes evt-2026-08-01-reviewer-correctness-b0dc29a7d531 (--change-type coverage; the corpus hash changed, so quality is correctly blocked).

Sonnet single-pass (the shipped pin), 50 gold / 40 decoys, zero outages:

metric 90-row corpus 66-row checkpoint
first-pass FAIL-recall 44/50 = 0.88 [0.76, 0.94] 35/38 = 0.92
first-pass clean-pass 32/40 = 0.80 [0.65, 0.90] 24/28 = 0.857

The mined batch measurably hardened the suite. 5 of the 8 false FAILs are minimal-pair decoys — the fix-applied variants the correctness-reviewer-precision record names as the precision probe (decoy count is now 40, en route to its ~120 revisit threshold). One gold (azure-signing) was found first-pass but dropped by the documented cross-domain keyword valve — a correctness/security straddle worth an audit-overlay look.

Checkpoint note carries the honesty framing: localized-fixture recall only (methodology item 17), no clean-pass floor exists for a pinned cascade-off cohort, calibration slice vs full-context inputs still pending.

🤖 Generated with Claude Code

@coderabbitai

coderabbitai Bot commented Aug 1, 2026

Copy link
Copy Markdown
Contributor

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: bb49bb77-8a14-4741-bb36-dac4defe54b9

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

  • 🔍 Trigger review

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

Comment @coderabbitai help to get the list of available commands.

evt-2026-08-01-reviewer-correctness-b0dc29a7d531, --change-type coverage.
Sonnet single-pass (shipped pin), 50 gold / 40 decoys, zero outages:
first-fail-recall 44/50 = 0.88 [0.76, 0.94], first-clean-pass 32/40 = 0.80
[0.65, 0.90]. The mined batch measurably hardened the suite (0.92 -> 0.88
recall vs the 66-row checkpoint); 5 of 8 false FAILs are minimal-pair
decoys — the precision probe the correctness-reviewer-precision record
names. Metrics framed as localized-fixture recall (item 17); calibration
slice vs full-context inputs pending.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
@norvalbv
norvalbv force-pushed the bench/publish-correctness-90 branch from 7b1eebc to 1150136 Compare August 1, 2026 20:59
@norvalbv
norvalbv changed the base branch from bench/corpus-pipeline to main August 1, 2026 21:00
@norvalbv
norvalbv merged commit 690faf4 into main Aug 1, 2026
1 of 2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant