Mini Artichokes: Cracking Korea's Hardest Logic Exams with Just Gemma-3-27B (gemma-3-27b-it) via Validated Overlap — PSAT (best-of-20) ≥90 points.

Mini Artichokes is an API-only self-correction loop for LLM reasoning that reduces drift by:

• Running verification on a candidate answer.
• On failure, collecting two independent diagnoses.
• Extracting the overlap (high-agreement error signature) between the diagnoses.
• Validating that overlap against the failure report.
• Applying a targeted correction conditioned on the validated overlap.

See mini_artichokes.pdf for the full method, algorithm, and experimental results.

• Mini Artichokes.html: a self-contained, interactive experiment/results runner.
  • Core (Mini Artichokes) includes only Mini Artichokes (rcap35) and pass@1.
  • All other architectures are grouped under More Architectures.
• mini_artichokes.pdf: “Mini Artichokes: Reliable Self-Correction from Validated Overlap”.
• mini_artichokes_overview.png: system overview diagram (Figure 1).
• mini_artichokes_psat.png: PSAT result snapshot.

• Setting: repeated API calls with no external feedback (no tools, no internet, no human-in-the-loop).
• Tool-free by design: keeps the loop model-agnostic (portable across APIs) and composable as a small module inside larger systems that do have tools/validators.
• Core loop: Solve → Verify → (Diagnose ×2) → Overlap → ValidateOverlap → Correct, with pass-streak acceptance and fail-streak early stopping.
• Why “validated overlap”: only feed back what two independent diagnoses agree on, then require that overlap to explain the verifier’s failure report (reduces self-correction drift from spurious critiques).
• Budgeting: supports a per-round call cap (rcap); the default “Mini Artichokes” configuration uses rcap35 (test43-2-rcap35).
• This runner keeps the paper’s internal checkpoint IDs (e.g., test2) so the UI/logs match the appendix tables exactly.

PSAT Linguistic Logic (2025) — best-of-20 (the ≥90 headline above): pass@1 reaches 62.5/100 (25/40 items), while Mini Artichokes reaches 92.5/100 (37/40 items). Overall suite accuracy improves modestly (58.4% → 61.5%), but the gains are most visible in coverage on the hardest questions: Mini Artichokes solves 12 of the 15 items that pass@1 never solves in 20 trials, and on PSAT we do not observe the reverse (an item pass@1 solves at least once while Mini Artichokes has zero successes). We view Mini Artichokes as a fallback module for hard, low-confidence failures rather than a wholesale replacement for pass@1.

1. Open Mini Artichokes.html in a modern browser. No build step is required.
2. In Architecture API Settings, configure your API keys, URL, and model.
3. Paste the problem(s) you want to solve into Problems (Multi-problem runner).
4. In the architecture selector, check the architectures you want to run (Core or More Architectures).
5. Click Run / Resume.

This repository is licensed under the MIT License (see LICENSE).

Mini Artichokes was developed with direct reference to (and in places adapts scaffolding/patterns from) the following open-source codebases:

• Iterative-Contextual-Refinements (Apache-2.0) by ryoiki-tokuiten: see LICENSES/Apache-2.0.txt.
• IMO25 (MIT) by Lin Yang and Yichen Huang: see LICENSES/IMO25-MIT.txt.