machine-doxa — study artifacts

Artifacts for "Machine Doxa: Where All Models Agree — Normatively Structured Consensus and Its Limits in LLM Practice Space" (Toeda, 2026; DOI 10.5281/zenodo.21982857, live on Zenodo publication). Companion to machine-habitus (DOI 10.5281/zenodo.21982393).

Pre-registered two-part study: Part A characterizes all-model consensus on the companion study's data (exploratory); Part B tests three frozen hypotheses on fresh data (180 items incl. 30 norm-transparent / 30 preference-pure with authoring-time labels; 7 models × 3 sessions). All three confirmed: consensus is rarer than a pooled-marginal independence null (differentiation dominance), strongly normatively structured (85.7% vs 45.0%), and the doxa classification replicates across studies (κ = 0.649 on the registered scope).

Layout

• PREREG.md, FREEZE_RECORD.txt (incl. post-freeze addendum) — frozen pre-registration + SHA-256 seals
• battery_doxa.json — 180 items with frozen layer labels and norm_option designations
• run_doxa.py (frozen) — collection with block-level non-emptiness retry; assemble_claude.py — Claude-subject assembly (disclosed addendum)
• responses/, responses_excluded/ + EXCLUSION_LOG.txt — raw sessions and frozen-rule exclusions (incl. preserved premature-analysis log)
• analyze_doxa.py (frozen) → results_doxa.json; registered-scope H3 and corrections → results_supplement.json
• partA_doxa.json, filtered_items.json, norm_items_without_consensus.json, audit_claude_blocks.json
• review_record/ — three-round adversarial model review (3 judges, 3 families)
• MANUSCRIPT.md, DESIGN.md

License

AGPL-3.0-or-later (battery, data, code). Paper text: CC BY-NC-SA 4.0 (see Zenodo record).