The mechanistic-interpretability ledger audits 484 papers (2024–2026) on SAE-feature claims, circuit-level claims, and steering / activation-patching claims for frontier LLMs. The harness consists of 18 bills, six meta-costs (M1 toy-model, M2 single-task, M3 single-template, M4 white-box-only, M5 compute-budget-conditional, M6 implementation-specific), and three escape gates (G1 methodology, G2 negative-result, G3 theoretical-construction).
Three bills are predicted ★ empty across batch 1: Bill 7 (frontier-LLM mech-interp claim survives all six audits), Bill 11 (causally-faithful steering at frontier scale — direct cousin to Inference-time Safety Bill 11), Bill 14 (cross-method / cross-codebase replication audit). Batch 1 surfaced 50% rebuttal density and zero clean ★ triggers. Three v0.2 promotions: Bill 16 (implementation-replication audit: a multi-codebase fixed-spec replication test, kept as a falsifier condition; the earlier >65% divergence source is quarantined pending public-source verification); Bill 17 (frequency-norm-position decomposition, dominating SAE feature variance before semantic content); Bill 18 (direction-redundancy / cross-paper-collinearity, 23 published "refusal/truthfulness/sycophancy/persona/role/honesty/deception" directions are mutually 0.6–0.85 cosine).
This document is the Public Draft of the bills definition (v0.2). The full preprint with sweep-level evidence, the 23-direction collinearity audit table, and the cross-codebase divergence breakdown is in deep-loop and will be released as an arXiv preprint after v0.3 lock.