worldexams-quality-gate
Use when validating or repairing generated WorldExams/SaberParaTodos bundle markdown before opening a PR. CJK/English leakage, evidenced ICFES axes, answer-letter rebalance, dual validators.
- License 许可
- MIT
- 所属仓库
- iberi22/skills
- 最近更新
- 2026-10-12
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- worldexams-quality-gate
- description
- Use when validating or repairing generated WorldExams/SaberParaTodos bundle markdown before opening a PR. CJK/English leakage, evidenced ICFES axes, answer-letter rebalance, dual validators.
03 SKILL.md 正文
WorldExams quality gate v2
Validates generated questions_data/**/weekly/*-MASTERY-bundle.md before it reaches a PR.
Calibrated against already-merged content: a gate that flags merged files is mis-specified, not the content.
Two validators, both must pass
node scripts/validate-bundles-v52.mjs <file.md> [more.md ...]
node saberparatodos/scripts/validate_content.js --country=colombia --grade=<N> # needs node_modules
WHERE EACH ONE ACTUALLY RUNS, verified 2026-10-01:
validate-bundles-v52.mjsruns in CI (.github/workflows/deploy-preview.yml), on the changed bundle files. It isnpm run validate.- It does NOT run in
.husky/pre-commit. The hook runs three things: the SWAL leak gate,scripts/husky-guard.mjs --staged, and the interlock that blocks unsafe generators. A bundle with bad feedback commits locally and is caught by CI, not by the hook. validate_content.jsis a separate older validator undersaberparatodos/. It is not wired into the hook or intodeploy-preview.yml.
So a green git commit does not mean the content passed. Run the validator
yourself before pushing, or wait for CI.
validate_content.js needs saberparatodos/node_modules. In a fresh worktree, symlink it from the
main checkout, run, then delete the symlink. To validate files that only exist in a worktree, copy
them into the main checkout, run, then delete the copies.
Defects the generator actually produces
| Defect | Detection | Fix |
|---|---|---|
| CJK injected into Spanish prose | grep -nP '[\x{4e00}-\x{9fff}\x{3040}-\x{30ff}\x{ac00}-\x{d7af}]' |
rewrite the phrase in Spanish |
| Cyrillic/Greek in words | grep -nP '[\x{0400}-\x{04ff}\x{0370}-\x{03ff}]' |
transliterate or rewrite |
English words in Spanish prose (signalling, el_noise) |
word list over slug-stripped text | replace with Spanish |
Duplicated **Contexto:** line |
count each field vs question count | drop the 2nd (mangled) copy |
A question missing **ICFES:** |
ICFES count != question count |
restore the evidenced axis |
| Answer-letter bias (7/12 on B, or zero D) | repo validator warning | rebalance (below) |
The CJK grep range in the cron spec is malformed: \x{ac00}-\xd7af errors with
"range out of order in character class". The working endpoint is \x{d7af}.
Always confirm a detector detects: append real CJK to a temp copy and grep it.
Accent check must skip slugs
Protocol requires ASCII kebab slugs and **ID:** values, so el-realismo-magico is CORRECT and
Realismo Mágico in prose is also correct. Strip slugs first, then look for the unaccented word:
re.sub(r"[a-z0-9]+(?:-[a-z0-9]+)+", " ", line) # remove slugs/ids
re.search(r"\b(magico|tambien|fisica|quimica|cancer|farmacos)\b", stripped)
Do NOT add real Spanish cognates to the English blocklist (defender, correct, answer are
valid Spanish and appear in merged content).
ICFES axes must be evidenced
An axis is valid only if it already appears in >=3 merged bundles of the SAME subject. Compute the
set from the subject dir rather than guessing. For lengua the evidenced trio is:
Competencia Lectora (Literal)Competencia Lectora (Inferencial)Competencia Lectora (Crítica)
Forbidden: the subject name as the axis, combined forms, and axes from another subject.
Difficulty ranges
validate_content.js only requires each header be in the ALLOWED SET
{D3-D4, D5-D6, D7-D8, D9-D10} - it does NOT enforce a fixed distribution. A 3/3/4/2 split is
tolerated in merged content. Enforce monotonic (non-decreasing) progression as an error; treat
distribution drift as a warning only.
Answer-letter rebalance
Rotate the correct option as a WHOLE LINE so its <!-- feedback: --> travels with it, then
relabel A)-D). Never rewrite option text. Target 3/3/3/3 in a 12-question bundle; move only from
over-represented letters to under-represented ones. Verify on a copy first: option text + feedback
sets must be identical per question before/after.
Static packs
Never hand-edit packs. Regenerate from the markdown:
node saberparatodos/scripts/generate-static-packs.js --all-weekly --changed-only
.husky/pre-push calls scripts/husky-guard.mjs --changed, which regenerates the
changed packs itself and then calls assertGeneratedPacksCommitted(). So a
commit that leaves a pack unstaged is refused at push time, not at commit time.
After generating by hand, verify each pack's correct_answer matches the
markdown [x] letter and that every options[].feedback is populated.
Commit hygiene
The pre-commit hook validates only STAGED bundles, so stage the verified subset explicitly rather
than git add questions_data/. A failing W31 in the tree must not block a good W27-W30 commit.
Related
- [[hermes-cron-worldexams-pipeline]] - the 30-min autonomous tick
04 同仓库更多 Skills
web-design-guidelines
暂无简介
blocked-page-recovery
Use when a fetch fails: 403/429, paywall, WAF, bot wall. Recovers the page via archives and provenance-checked fallbacks.
studio-coach
暂无简介
src-generator
暂无简介
sqlite-pro
暂无简介
rigor-evaluator
Use when you need a quick heuristic rigor score (10 criteria, 0-2 each) for the features in a .gitcore/features.json file, to decide which features are ready to ship.