Rule Share Is Not the Lever (Wolfberg research note, 29 October 2026): data and code
https://wolfberg.ai/papers/rule-share-is-not-the-lever/

REGISTRATION.md      design, measures, predictions and decision rules, written before any outcome
AMENDMENT-1.md       the one change (bfloat16 on a local GPU), written before any model outcome
RESULTS.md           all estimates and verdicts
armM.py              the model-arm run script. Usage: python armM.py <path to Phi-3.5-mini-instruct> <out.jsonl>
                     env: DEV=cuda|cpu, DT=bf16|fp32, FORCE_A=1 (rule A grid), ONLY_K0=1 (no-history cells only).
                     Needs torch and transformers 4.57.1; set TORCHINDUCTOR_CACHE_DIR if your OS user name is unavailable.
armM_bf16.jsonl      one row per prompt (primary run): condition, k, tokens, rule share (sink excluded and included),
                     per-layer shares, the reply, rule-breaking (strict and lenient), sum correct, timings
armM_bf16_ruleA_exploratory.jsonl   the rule A grid (post hoc)
armM_fp32cpu_k0.jsonl               float32 CPU rerun of the no-history cells (precision check)
armM_cells_*.csv     cell means; armM_by_layer_ruleB.csv per-layer shares for key cells
armR_boundaries_K5_pbflag.csv       record arm, one row per compaction boundary (counts only; sessions relabeled S01 to S21)
Hypothesis and H1 design: Rock Lambros, The Dual-Failure Multiplier (CxBOM, cxbom.org).
