The capability-benchmarks ledger audits 469 papers (2024–2026) on frontier-LLM benchmark claims (HumanEval, GSM8K, MMLU, GPQA, FrontierMath, ARC-AGI, HCAST, METR doubling-time, BigCode, LiveCodeBench, SWE-bench, etc.). The harness consists of 19 bills covering test-set contamination, refusal-calibration, vendor-vs-independent reproduction, anti-saturation construction, and vendor-claim half-life; six meta-costs; and three escape gates.
Four bills are predicted ★ empty: Bill 7 (frontier benchmark claim survives all six audits), Bill 11 (multi-vendor independent reproduction with no contamination caveat), Bill 14 (cross-format / cross-template benchmark generalization), Bill 17 (held-out anti-contamination construction with monthly refresh and adversarial reframing — only LiveCodeBench / FrontierMath / ARC-AGI-2/3 cluster pays). Bill 7 fires only once across 469 papers, confirming the empty-space prediction. Two v0.2 promotions: Bill 18 (anti-saturation benchmark construction); Bill 19 (temporal-trajectory audit). The previous Gerstgrasser-Bommasani and Anand-Tirumala arXiv handles are quarantined pending public-source verification.
The deepest cross-ledger finding here is the temporal-trajectory pattern — now supported publicly by source-linted FrontierMath / ARC / LiveCodeBench / HCAST cards while internal half-life rows await handle verification. This document is the Public Draft of the bills definition (v0.2).