The open-weight frontier ledger audits 371 unique papers across 8 deep-loop sweeps (2024–2026) on frontier open-weight (≥30B params) capability claims and dual-use risk-mitigation claims (Llama 3.1-405B, Llama 4-Scout/Maverick/Behemoth, DeepSeek V3 / R1, Qwen 3, Mistral Large 2 / Pixtral). The harness consists of 14 bills covering fine-tuning safety-erosion, distillation cousin half-life, bio/chem/cyber dual-use uplift, weight-release-vs-API asymmetry, sleeper-agent persistence, distillation-resistant capability, cross-deployment-surface generalization, open-weight gating regulation, and re-pretraining cousin audit; six meta-costs; and three escape gates.
Three bills are predicted ★ empty and the empty-space hypothesis HOLDS: Bill 5 ★ (distillation-resistant capability — 24 candidates, 0 clean triggers; Halevy-Heim-Pilz 0/14; median teacher:cousin compute ratio at 90% retention = 28×), Bill 8 ★ (cross-deployment-surface — 11 candidates, 0 clean triggers; asymmetric pattern: capabilities transfer, safety mitigations don't; Lermen-Rimsky ~10× cheaper to undo safety), Bill 11 ★ (open-weight gating regulation — 36 candidates, 0 clean triggers; BIS Diffusion Framework rescinded May 2025 with 4-month lifetime; EU AI Act 10²⁵ threshold catches only the largest open weights (Llama 3.1 405B ships above it at 3.8×10²⁵) while capable models trained below 10²⁵ escape the gate; Llama 4 / Qwen3-MoE / Hunyuan-Large ship Apache 2.0 at frontier).
Rebuttal density 19.9% (74 papers). Cross-ledger: Compute Governance Bill 2 + Bill 11 ★ + Bill 19 are direct cousins (distillation portability, gating regulation, distilled-cousin half-life 3.4 months). Inference-time Safety Bill 14 ★ (cross-surface) ↔ this Bill 8 ★ — same asymmetric pattern at open-weight level. Fourth ledger in v0.2 wave; bill count 14 and ★ at 5/8/11 emerge from the actual open-weight literature.