Skip to content

Validator prompt version; regression runs and re-check data kept per version - #910

Merged
mircealungu merged 3 commits into
masterfrom
validator-form-rule
Oct 25, 2026
Merged

mircealungu merged 3 commits into
masterfrom
validator-form-rule

Conversation

@mircealungu

@mircealungu mircealungu commented Oct 25, 2026 •

Copy link
Copy Markdown
Member

No prompt change. Versions the validator prompt and keeps its evaluation data in the repo.

  • PROMPT_VERSION in prompts/translation_validator.py: translation-validator-v1 is the deployed prompt (Validator: the meaning must fit the sentence it is shown in #900). Bump it on every prompt change.
  • tools/run_validator_regression.py --save writes each run to regression/results/translation_validator/<version>--<date>.txt, with a README table of the versions.
  • tools/recheck_validator_fixes.py says which version judged.
  • The audio audit's checker_id includes the version, so code changes around the prompt also make old reports unusable for --deprecate.
  • Evaluation data: regression/recheck/translation_validator--2026-10-25--top300.csv (the first re-check of past fixes: 64 disagreements in the top 300) and 9 new regression cases from it.

Tried and not shipped (runs kept): v2–v4 tuned how strictly the card must match the translation's own grammar (Dutch echt vs echte). The best, v3, scored 48/58 against v1's 46/58, which is within run-to-run noise, for a longer prompt with exceptions. Instead the plan is to keep the looked-up form on the bookmark and the learning unit on the user word (docs future-work/bookmark-and-learning-unit.md). The adjective cases stay in the set as known v1 misses.

🤖 Generated with Claude Code

mircealungu and others added 2 commits October 25, 2026 20:22
…er version

PROMPT_VERSION names the prompt (v1 = api#900). The regression runner saves
its report with --save, and the re-check tool says which version judged. The
audio audit's checker_id includes the version. Keeps the 25 Oct re-check of
the top 300 past fixes and the v1 baseline.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ws the word

The 25 Oct re-check of past fixes found v1 demanding the translation's own
agreement: Dutch "echt" -> "echte" because Dutch writes "een echte
uitdaging", though the learner sees the French sentence. v3 exempts adjective
endings and articles only the translation needs; tense, number and person
still follow the word (Nutzer = gebruikers). 48/58 against v1's 46/58; v2 and
v4, wider exemptions, were more lenient and scored lower. New cases from the
re-check in the regression set.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@github-actions

Copy link
Copy Markdown

ArchLens - No architecturally relevant changes to the existing views

… runs and the cases

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@mircealungu mircealungu changed the title Validator v3: an adjective keeps its plain form; prompt versions with saved regression runs Validator prompt version; regression runs and re-check data kept per version Oct 25, 2026
@mircealungu
mircealungu merged commit 16c43c9 into master Oct 25, 2026
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