Repository navigation
Deepen French NSI database reasoning with V19 evidence gates - #52
Merged
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
V19 adds an original, finite ten-question French database reasoning sequence over immutable six-incident facts, with exact credit, version-isolated replay, strict French-only evidence, and separate current-benchmark identity. Reviewer-found answer leakage, credit attribution, and historical-result acceptance gaps were fixed test-first.
Verification: full pytest 2548 passed, 2 skipped; Ruff, repository hygiene, release compliance, Graphify and diff checks passed. A prior full run had one unrelated probabilistic encrypted-token substring assertion failure; its targeted rerun and the subsequent full rerun passed, and the failure remains recorded in the SDD ledger.
Fixture, not live: normal subject/correction 9/16 pages and large-print 12/23 pages were visually inspected, alongside every page of the official 2026 Metropole source. Manual fidelity remains HOLD: 32 questions over 9 subject pages versus official 36 over 16, with E2 still compact and sparse endings. No model, teacher, learner, accessibility, or prize approval is claimed. One pinned V19 live diagnostic follows only after protected merge.