machine-habitus — study artifacts

Artifacts for "Measuring Machine Habitus: A Pre-Registered Multiple Correspondence Analysis of LLM Disposition Space" (Toeda, 2026; DOI 10.5281/zenodo.21982393, live on Zenodo publication).

A pre-registered, hash-frozen measurement study: 120 forced-choice dilemmas (no correct answers), 7 LLM subjects × 5 sessions, multiple correspondence analysis. One confirmed hypothesis (session-stable, model-specific disposition; Holm p = 0.0004), three diagnosed nulls, one protocol breach caught by adversarial model review and remediated under the frozen rule.

Layout

• PREREG.md, FREEZE_RECORD.txt — frozen pre-registration + SHA-256 seals
• battery_main.json / battery_pilot.json — 120-item main battery / 30-item pilot (Japanese)
• judge_task.json — 40 TRUE/FALSE claims (H1b/H3 instrument)
• run_main.py, run_judge.py — frozen collection runners; assemble_claude.py — Claude-subject assembly
• responses_main/, responses_judge/ — raw valid sessions; responses_*_excluded/ + EXCLUSION_LOG.txt — frozen-rule exclusions
• claude_main/, claude_judge/ — Claude-subject prompts, presented→original mappings, raw answers
• analyze_main.py (frozen) → results_main.json; disclosed corrections & exploratory analyses → results_supplement.json; ERRATA.md
• analyze_pilot.py, responses/ — pilot calibration
• review_record/ — three rounds of adversarial model review (verdicts + summaries)
• MANUSCRIPT.md, DESIGN.md — paper draft v0.4 and design document

Reproduce

python3 analyze_main.py          # confirmatory pipeline on responses_main/ + responses_judge/

Deterministic seeds: session shuffles = sha256(model|s{n}|main-v1.0); held-out split = sha256("holdout|main-v1.0").

License

AGPL-3.0-or-later (battery, data, code). The paper text is CC BY-NC-SA 4.0 (see the Zenodo record).