RedJudge v1.1.0 adds the learning mode — an evidence-aware adversarial review of a learner's mastery of a knowledge point. Its failure target is false mastery (能背公式但换问法就崩): the learner repeats the standard answer but breaks on counter-examples, edge cases, and transfer problems.
What's new
Learning mode (new object type)
- 4-level mastery verdict bound to weighted score: 已掌握 / 部分掌握 / 表面掌握 / 未掌握 (mastered / partial / shaky / not_learned).
已掌握is intentionally hard to earn: weighted total ≥ 75 AND 反混淆 ≥ 7 AND no dimension below 6.- Default starting assumption is
表面掌握, not "probably fine". Verification is required to upgrade.
5-dimension mastery rubric
| Dimension | Weight | What it checks |
|---|---|---|
| 概念清晰度 | 25% | Can the learner state the concept in own words and name its preconditions? |
| 边界精度 | 20% | Can the learner name where the concept stops being valid? |
| 程序熟练度 | 20% | Can the learner execute the standard procedure on a fresh problem? |
| 迁移 | 20% | Can the learner apply the concept to a differently phrased problem? |
| 反混淆 | 15% | Can the learner name an adjacent concept and articulate the difference? |
反混淆 (Anti-Confusion) is the core dimension — false mastery hides here. The learner must distinguish the target concept from at least one adjacent one. If they cannot, the verdict cap is 部分掌握.
Mandatory output blocks (non-negotiable even in quick mode)
- 🎯 薄弱点 — concrete sub-skills or sub-steps where the learner failed, with evidence (which problem, which mistake).
- 🔀 混淆点 — adjacent concepts the learner appears to conflate with this one, with the specific mis-pairing and the discriminator that would break the tie.
- 📋 重学补丁 — concrete exercises or checks for re-study. Not "review chapter X" — a specific re-do.
Diagnostic lenses (Multi-Perspective Review)
Replaces stakeholder roles (target user, investor, editor) with: 严苛考官, 前置知识核查, 相邻概念对比者, 迁移出题人, 课标对齐.
Chinese-only naming
All learning-mode verdict names, dimension names, block names, role names, and template headers are now Chinese. The new references/localization-rules.md codifies this convention for Chinese-facing artifacts. English terms preserved only where they are established RedJudge vocabulary or trigger tokens: Type: learning, Red Scan, Evidence Boundary, weighted total, continue / revise / abandon (legacy 3-tier).
New artifacts
references/learning-dimensions.md— full rubric + scoring anchors + verdict mapping + anti-false-mastery rules.references/learning-output-format.md— full template + quick template + self-check.references/localization-rules.md— Chinese-only naming convention.examples/learning-review.md— worked example grading a learner's self-claim of mastering the quadratic formula (verdict: 表面掌握).
Trigger phrases
RedJudge learning：判一下我对 X 掌握得怎么样批改一下我做的这道题我这样算掌握 X 了吗RedJudge learning strict(5 failure modes, ≥ 2 confusion points required)RedJudge learning quick(skip multi-perspective and value confirmation; keep 薄弱点 / 混淆点 / 重学补丁)
Validation
python scripts/check-redjudge-evals.py: PASS (extended to require learning references, learning example, four-level verdict term, and 薄弱点/混淆点 terms).- English-leakage audit on 5 learning artifacts: clean (no bilingual aliases, no English template headers).
- GitHub Actions
Validate RedJudgeworkflow: ✅ success on commit55e4158. - New regression prompts added to
evals/evals.json(now 8 total):learning-shaky-false-mastery,learning-mastered-real-mastery,learning-not-learned-empty-evidence.
Scope boundary
RedJudge learning mode grades and diagnoses only — it does not write lessons, examples, or step-by-step tutorials. For fresh teaching material, hand off to a teaching skill.
Links
- Changelog: https://github.com/gaoyechen/redjudge/blob/main/CHANGELOG.md
- Repository: https://github.com/gaoyechen/redjudge
- Install:
npx skills add gaoyechen/redjudge