commit_tree.txt

v1.4 20260324 1625 — Phase C frontier models: Claude-3.7-Sonnet (#1) + GPT-4o (#10)
    - Claude-3.7-Sonnet: MEI=0.774 SR=56.7% HRR=87.5% d=0.554 (NEW #1)
    - GPT-4o: MEI=0.315 SR=6.7% HRR=35.3% d=1.917 (NEW #10, below GPT-4o-mini)
    - 10 models x 60 trials = 600 total; analysis_final2.json rebuilt k=10
    - Update hallumaze_final.html leaderboard + abstract + findings
    - Update paper draft Section 4: new results, F3 frontier inversion
    - Add experiment_results/or_phaseC.json

v1.3.1 20260324 — MARL v2 GLM-4.7 extension + paper/docs update
    - MARL v2 GLM-4.7: MEI=0.694 (+12.9% vs GLM baseline 0.615)
    - Add run_marl5stage_v2_glm.py + marl5stage_v2_glm.json
    - Paper draft: add Section 6.2 MARL experiment findings
    - Update docs/hf_post.md, DOC_INDEX.md

v1.3 20260324 — MARL 5-stage experiment (v1 + v2 with fixes)
    - MARL v1: naive 5-stage MEI=0.548 (-7.6% vs baseline)
    - MARL v2: PathValidator+ConditionalPipeline+MazeReinjection MEI=0.803 (+35.4%)
    - Add HF community post draft (docs/hf_post.md)
    - Add research docs: fix-solutions, ecological-validity

v1.1 20260323 — Add core library, fix statistics, HF Space ready

- Add files/hallumaze.py (core benchmark library, 951 lines)
- Add requirements.txt
- Add index.html for HuggingFace Space (self-contained, HF YAML frontmatter)
- Fix anonymous GitHub URLs → jaytoone/HalluMaze
- Fix statistics: Glass's delta + one-sample Wilcoxon (corrected d=0.8-2.1)
- Add docs/research/ (paper draft + extension TODO analysis)
- Add experiment_results/ full dataset (20 JSON files)
- Fix README: pip install requirements.txt, reproducibility notes