Skip to content

Calibrate the matcher on authentic pairs and reject reused indices - #46

Merged
MHohlios merged 1 commit into
mainfrom
feature/matcher-calibration
Oct 2, 2026
Merged

MHohlios merged 1 commit into
mainfrom
feature/matcher-calibration

Conversation

@MHohlios

@MHohlios MHohlios commented Oct 2, 2026

Copy link
Copy Markdown
Contributor

The matcher had only a five-pair spot-check, and its duplicate-index validation compared a value with itself. Add 30 hand-labeled authentic gold/candidate pairs, a repeatable calibration runner, and an isolated Codex matcher backend with explicit model and effort settings. Reject reused, fractional, and out-of-range indices and surface provider errors.

The Codex gpt-6.1-sol / medium run agreed on 30/30 pairs with 0/15 trap false matches. This validates only that configuration; historical Claude measurements remain uncalibrated at scale.

Validation: four offline matcher tests, five portability tests, markdown lint, and git diff --check pass.

Fixes #16.

Signed-off-by: Michael Hohlios <michael@vyttle.com>
@MHohlios
MHohlios force-pushed the feature/matcher-calibration branch from e249ce8 to 8df543c Compare October 2, 2026 19:24
@MHohlios
MHohlios merged commit 50b2980 into main Oct 2, 2026
2 checks passed
@MHohlios
MHohlios deleted the feature/matcher-calibration branch October 2, 2026 19:29
MHohlios added a commit that referenced this pull request Oct 2, 2026
The 15 zero-correctness-recall CRB cases contain pre-existing defects,
deliberate behavior, incorrect API assumptions, and mislabeled cleanup.
Add pinned-source rulings to all 35 labels, preserve imported
text/categories, and exclude ruled-out labels from matching and the
correctness denominator without renumbering recorded matches.

The historical re-judge preserves the original outputs: 9 eligible
correctness defects, 17 policy exclusions, 9 wrong or underspecified
labels. One eligible defect was already reported but tiered as
efficiency. Entirely excluded cases stay unscored rather than becoming
expected-clean. Pinned Keycloak callers and feature-version selection
exclude two initially accepted labels; recovery-credential ID loss
remains eligible through the reachable account deletion path.

Validation: four gold-judge tests, four matcher tests, markdown lint,
git diff --check, and source-based reproductions of partial-cache event
mispairing and integer-tuple hash stability pass. This PR is stacked on
#46 and does not claim current-release recall.

Fixes #13.

---------

Signed-off-by: Michael Hohlios <michael@vyttle.com>
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.

Matcher calibration at scale (the ~30-pair sample owed since add-reviso-review task 5.5)

1 participant