← Evidence index · 中文 · Main post
Each section: Claim · Why · Evidence · Analogue · Would update if · Conf (H/M/L).
Parent: node5 open weights tamper
P = 0.88 — No federal open-weight ban by 2028
Claim: US will not criminalize or ban open release of frontier-scale weights by 2028.
Why: Meta/LMSYS ecosystem, DC “innovation + China competition” frame, Trump EO 14409 pro-deployment; Fable ban targets deemed export to adversaries, not open release generally. Cruz coalition killed state regulation ban 99–1 — federal ban on open weights even less likely than training cap.
Evidence:
- (internal note) §1.2–1.3
- (internal note) — export control ≠ open-weight prohibition
- Meta Llama 4 open release strategy (2025–26 industry pattern)
- (internal note) — China open-weights PR (DeepSeek)
Analogue: Encryption export controls (1990s) — narrowed, never banned open crypto research.
Would update if: Bipartisan bill introduced banning >X param open release with industry split (not just Anthropic/Google).
Conf: H
P = 0.55 — Voluntary / regulatory model documentation (EU AI Act GPAI)
Claim: Major open models face documentation + some eval duties under EU AI Act / codes of practice by 2028.
Why: EU AI Act applies to GPAI providers including open; Code of Practice 2025–26; not the same as banning weights.
Evidence:
- (internal note) EU enforcement rows
- EU AI Act GPAI obligations (2024–25 framework)
Would update if: EU explicitly exempts open weights from all GPAI duties.
Conf: M
P = 0.35 — Fable-class export control on specific open models
Claim: US uses EAR/deemed-export to restrict access to specific frontier models (not ban training open).
Why: Already happened Jun 2026 Fable/Mythos — precedent for BIS acting on model access. P<0.5 because narrow, model-specific, not category ban.
Evidence:
- (internal note)
- node3 Trigger E2
Conf: M–H (precedent exists)
P = 0.15 — Structural safety–capability coupling at frontier
Claim: ≥1 major lab ships frontier model where removing safety directions measurably hurts MMLU/agent benchmarks, at scale.
Why: Your Qwen HarmBench work shows one Pareto point (Type-B LoRA) — not yet frontier-wide. Incentives weak: open-weight race rewards decoupled safety. CAIS RepE line is research not product default.
Evidence:
academia_application/cais_meeting_july2026/05_ai_safety_views.md— Type-B LoRAcode/typebits/— steering separable from capability on open models- Arditi et al., “Refusal in Language Models Is Mediated by a Single Direction”
Would update if: You or CAIS publish coupling at 70B+ with replication by second lab.
Conf: L–M
P = 0.05 — Mandatory tamper-resistant training standard (federal)
Claim: Binding US rule that frontier training must pass tamper-resistance eval.
Why: No legislative coalition; harder than CBRN screening; Meta opposes. Lower than pause P(<0.05).
Evidence:
- (internal note) — pause dead
- Node 1 federal cap <0.05
Conf: H (for “very unlikely”)
P = 0.18 — T5-B: Open-weight accelerates CBRN/bio tail
Claim: Conditional on open-weight modal, ~18% of bio-extinction tail mass routes through tampered open models (jailbreak + local fine-tune).
Why: WMDP unlearning bypass easier on weights you own; correlates Node 2 TAIL-C (arbitrage). Not independent.
Evidence:
- (internal note) §2.1 jailbreak
- (internal note) TAIL-C
- CUT/WMDP unlearning literature
Would update if: Closed API-only models show higher misuse rate than open tampered models in incident data.
Conf: L–M
P = 0.12 — T5-A: Post-incident weight licensing
Claim: ~12% chance high-profile attack → EU/US licensing regime for releases above threshold.
Why: Smaller than bio screening coalition; tech industry split (Meta vs Anthropic). EU more likely than US.
Evidence:
- SB 1047 veto — hard caps fail
- EU Biotech Act screening path more politically viable than AI weight ban
- node1 T3 vetting law 0.08 — similar magnitude
Conf: L–M