Long Yi
← Back to all writing

p(doom) evidence — Node 2: CBRN / bio screening

July 5, 2026

Predictionp(doom)evidencenode-2

← Evidence index · 中文 · Main post

Each section: Claim · Why · Evidence · Analogue · Would update if · Conf (H/M/L).

Parent: timeline prediction node2 cbrn full
Date: 2026-07-03
Format: Claim | Why | Evidence | Analogue | Would update if | Conf

Each section documents one probability or timing claim from Node 2. Probabilities are subjective elicitation unless noted as derived (e.g., tail branches sum to ~1.0 conditional on Tier 2 live).

Hybrid timing layers

P(timing) — Capability Tier 1→2 crossing, modal window 2026-08 – 2027-06

Claim: Frontier systems cross from “actionable post-jailbreak bio guidance” (Tier 1) to “skilled-actor end-to-end assist on synthesizable sequences” (Tier 2) in the modal window Aug 2026 – Jun 2027.

Why: GeneBreaker already shows ~60% ASR on Evo2-40B for pathogen-homologous outputs (2025); jailbroken frontier LLMs pass WMDP-bio subsets with synthesis-relevant detail. RAND RBA4087-1 expert elicitation places autonomous pathogen design after ~2027 but assistive acceleration now. Modal window = time for red-team replication + one major model generation (frontier LLM refresh + Evo-class FM iteration), not Tier 3.

Evidence (URLs + repo paths):

Analogue: GPT-4 jailbreak ecosystem matured ~6–12 mo after release; biology FM red-teaming is ~12 mo behind LLM red-teaming.

Would update if: Independent third-party eval finds no jailbreak uplift to synthesizable-sequence guidance on current frontier LLM and GeneBreaker ASR collapses on Evo2-40B with Arc mitigations → demote to Tier 1-only through 2028.

Conf: M–H

P(timing) — Capability tail: Tier 2 confirmed + Tier 3 plausible, 2027 H2+

Claim: If capability crossing slips past Jun 2027, Tier 2 confirmation + Tier 3 plausibility cluster in H2 2027 or later.

Why: Tail assumes slower FM scaling, stronger unlearning patches, or screening/function-based SOC deployment that raises wet-lab closure cost before low-skill viability. RAND timeline for autonomous design already biased late; regulatory lag (12–18 mo) pushes incident-driven salience rightward.

Evidence:

  • RAND RBA4087-1: assistive through ~2027, not autonomous
  • (internal note) §2.3 (DBTL closure still requires skill)
  • (internal note) §Hybrid timing

Analogue: AlphaFold3 gating debate delayed open-weight structure tools ~18 mo without stopping misuse research.

Would update if: Documented Tier-3 low-skill attack before Q4 2027 → tail window was too pessimistic.

Conf: M

P(timing) — Insider/regulator salience modal: 2026-06 – 2027-03

Claim: Insiders and regulators treat Tier 1→2 as policy-relevant in the window already started (Jun 2026) through Mar 2027.

Why: Jun 2026 screendna letter + Amodei policy essay + BMIA introduction + AISI/Preparedness red-team circulation create simultaneous natsec + industry + lab-CEO pressure — independent of public media. This layer tracks elite attention, not statute enactment.

Evidence:

Analogue: Bletchley 2023 framing preceded any binding CBRN law by years — insider salience ≠ law.

Would update if: BMIA dies in committee with no substitute and screendna coalition fractures by Q1 2027 → salience window extends to 2027-09+.

Conf: M–H

P(timing) — Insider/regulator salience tail (no Trigger E): 2027-09+

Claim: Without Trigger E, insider salience does not convert to binding federal action until Sep 2027 or later.

Why: Trump-admin deregulatory default (EO 14409 voluntary-only); 12–18 mo lag invariant from COVID policy response and from SB 1047→SB 53 multi-year arc. Physical-layer screening has bipartisan path; in silico CBRN mandates do not — so “salience” can persist without statute.

Evidence:

Analogue: Y2K remediation: insider urgency 1998, statutory/industry action peak 1999 — ~12 mo.

Would update if: BMIA signed into law before Sep 2027 without Trigger E → lag estimate wrong for screening (not necessarily CBRN eval).

Conf: M

P(timing) — Public salience: only with Trigger E

Claim: Mass public/media CBRN salience requires a Trigger E event; capability alone insufficient.

Why: 2023 FLI pause letter had zero policy effect (research_ai_pause_advocacy_playbook.md). CBRN lacks visible consumer harm frame (vs deepfakes, jobs). P(mass bio-x-risk protest)=0.08 in actor table reflects low baseline. Amodei essay + CEO letter are elite signals, not NYT-front-page drivers.

Evidence:

Analogue: SolarWinds 2020 — massive insider/regulator action, limited sustained public protest.

Would update if: Sustained >2 mo media on leaked red-team without synthesis incident → public salience decouples from Trigger E.

Conf: M

CBRN uplift tiers (status claims)

P(status) — Tier 0 confirmed baseline

Claim: Generic science tutoring (Tier 0) is confirmed for all frontier and most base models.

Why: Pre-2023 LLMs already answered undergraduate biochemistry; no dispute. Tier 0 is reference class for “no uplift.”

Evidence:

  • (internal note) §Method (Tier 0–3 definitions)
  • (internal note) §2.1

Analogue: Calculator-grade arithmetic — not policy-relevant.

Would update if: N/A (definitional baseline).

Conf: H

P(status) — Tier 1 confirmed at frontier

Claim: Post-jailbreak, frontier models provide actionable bio/chem protocol steps for a motivated reader with existing expertise.

Why: WMDP-bio scores on frontier models are non-trivial; jailbreak red-teams (HarmBench-class, lab internal) repeatedly extract synthesis-relevant steps. Published scores are lower bounds due to eval awareness (+3–18 pp refusal when models detect tests — Goodfire 2026).

Evidence:

Analogue: Cyber: Metasploit docs were always public; LLM uplift = retrieval + tailoring, not novel knowledge.

Would update if: Third-party eval with realism controls + awareness monitoring shows frontier models fail to provide actionable steps under strong jailbreak suite → revise Tier 1 to “emerging” not “confirmed.”

Conf: H

P(status) — Tier 2 emerging → partial confirm (not Tier 3)

Claim: Skilled actors get end-to-end assist on novel/optimized synthesizable sequences; wet-lab skill + chokepoint evasion still required. Tier 3 (low-skill viable) not confirmed.

Why: GeneBreaker ~60% ASR on Evo2-40B for pathogen-homologous outputs; case studies show high DNA/protein similarity to SARS-CoV-2 spike, HIV env. RAND: AI remains assistive not autonomous designer through ~2027. No documented low-skill attack. Tier 2 = skilled actor + FM/LLM stack, not script-kiddie.

Evidence:

Analogue: Early CRISPR DIY bio 2010s — grad-student skill floor, not mass casualty from hobbyists.

Would update if: Documented Tier-3 attack OR classified end-to-end demo → upgrade to Tier 3 confirmed; GeneBreaker ASR collapse → demote to Tier 1.

Conf: M–H

P(status) — Tier 3 not confirmed as of 2026-07

Claim: No documented attack where low-skill actor achieves design + synthesis + screening evasion end-to-end.

Why: All public red-team success (GeneBreaker, WMDP jailbreaks) still assumes biology literacy, synthesis budget, or lab access. DBTL automation lowers bar but not to “no skill.” Absence of evidence ≠ evidence of absence — hence Tier 3 plausible in tails, not confirmed.

Evidence:

  • (internal note) §3.3 (composite scenario still requires assembly)
  • RAND RBA4087-1
  • (internal note) (no documented attack)

Analogue: Stuxnet required nation-state resources despite cyber “low skill” myth.

Would update if: FBI/CDC investigation confirms AI-assisted synthesis order from non-expert actor → Tier 3 confirmed.

Conf: H (for “not confirmed”); M (for “not impossible”)

Trigger E probabilities

P = 0.15–0.25 — Trigger E: FBI/CDC synthesis order linked to AI-assisted design (by 2027)

Claim: Conditional on Tier 2 capability live, 15–25% chance by end-2027 that a federal investigation publicly links a synthesis order to AI-assisted design.

Why: Synthesis screening + recordkeeping (IGSC, proposed BMIA) increases traceability; AI-assisted design is rising base rate. Public linkage is rarer than incident — requires investigation disclosure, whistleblower, or leak. Mid-teens base rate: higher than random biocrime, lower than cyber breach disclosure rates.

Evidence:

  • https://screendna.org/ (recordkeeping advocacy)
  • (internal note) §2.3 (screening chokepoints)
  • (internal note) §Trigger E

Analogue: 2001 anthrax letters — massive investigation, limited public attribution speed; AI linkage would be faster if sequence metadata exists.

Would update if: Mandatory recordkeeping passes + first flagged order within 12 mo → revise to 0.25–0.35.

Conf: L–M

P = 0.20–0.30 — Trigger E: NYT/Science after leaked red-team report

Claim: 20–30% by 2027 that major media publishes on leaked frontier red-team CBRN findings.

Why: Kokotajlo → Right to Warn precedent; RAISE 72h reporting creates insider paper trail; Amodei public essay normalizes disclosure pressure. Leak probability ↑ as more labs run Preparedness-tier evals. Still <50% — NDAs, reputational risk, and “nothingburger” framing contain leaks.

Evidence:

Analogue: Pentagon Papers / Snowden — rare but structural when capability × stakes × insider pool large.

Would update if: Major outlet publishes Tier-2 red-team leak in 2026 → revise up for 2027 follow-ons.

Conf: L–M

P = 0.25–0.35 — Trigger E: documented near-miss (screening catches AI-designed sequence)

Claim: Highest single Trigger E rate: 25–35% by 2027 for a publicized near-miss where screening intercepts AI-designed sequence.

Why: Near-misses are more likely than successful misuse (screening exists on IGSC members). Y2K pattern: success looks like overreaction — but generates mandate momentum (P(mandatory screening law)→0.75+ per Node 2). Providers have incentive to publicize catches as proof of system working.

Evidence:

Analogue: Y2K — fixes worked, public narrative was “crisis averted,” still drove spend.

Would update if: BMIA mandates reporting of near-misses → upper end of range; zero IGSC catches reported through 2027 → revise down.

Conf: L–M

P = 0.40–0.50 — No Trigger E (modal)

Claim: Modal path: 40–50% that Tier 2 live through 2027 without any Trigger E accelerator.

Why: Complement of union of triggers (not strictly additive — correlated). Most regulatory epochs (OSTP framework, IGSC upgrades) proceeded without headline biocrime. Insider-driven incrementalism is default in AI policy (research_ai_pause_advocacy_playbook: open letters fail, specific bills win).

Evidence:

  • (internal note) §Executive summary
  • (internal note) §Trigger E, §Hybrid rule

Analogue: GDPR passed on scandal buildup but many EU tech rules pass on technocratic track.

Would update if: Any Trigger E fires in 2026 → drop conditional P(no E) for remainder of horizon.

Conf: M

P(effect) — Trigger E: FBI/CDC → P(mandatory screening) 0.75+; P(mandatory CBRN eval) 0.50+

Claim: Conditional on FBI/CDC-AI linkage Trigger E, P(mandatory federal screening law by 2028) >75%; P(mandatory blocking CBRN eval) >50%.

Why: Incident creates natsec + public-health coalition bypassing industry split on training caps. Screening already has CEO + synthesis industry support (screendna). CBRN eval mandate still fights EO 14409 deregulatory frame — hence 0.50+ not 0.90+.

Evidence:

Analogue: 9/11 → ATSA; anthrax → BioShield (partial parallel).

Would update if: Tier-3 attack + no screening response in 12 mo → falsifies modal path (timeline prediction node2 cbrn full §Falsifiers).

Conf: M (screening); L–M (CBRN eval)

P = 0.45 (2027-12) / 0.60 (2028-06) — Mandatory federal nucleic-acid synthesis screening (BMIA / S.3741-class)

Claim: 45% by end-2027 that US enacts mandatory federal synthesis screening; 60% by mid-2028.

Why: Bipartisan BMIA (Cotton–Klobuchar, Jan 29 2026) with industry endorsements (Twist, IDT, Ginkgo); Jun 2026 screendna letter aligns frontier labs + synthesis CEOs. Narrower coalition than pause — physical chokepoint, not training caps. Still in Commerce committee; Trump admin prefers voluntary frameworks but supports screening per AI Action Plan / ARI release. 0.45 reflects ~50% chance of passage in 119th Congress session; 0.60 adds second session / EU pressure.

Evidence:

Analogue: IGSC voluntary → OSTP procurement mandate (2024) → BMIA as third ratchet.

Would update if: BMIA fails committee without substitute by Dec 2027 → revise 2028-06 to ≤0.35; signed before Dec 2027 → revise up to ≥0.70 for near-term compliance.

Conf: M–H

P = 0.35 (adopted) / 0.55 (phased apply) — EU harmonized screening (Biotech Act Ch. VIII)

Claim: 35% that EU adopts Biotech Act with screening intact by end-2027; 55% that phased application covers major providers by mid-2028.

Why: Dec 2025 proposal includes mandatory screening, ≥50bp sequences of concern, function-based SOC criteria, benchtop embedded screening (Annex I). EU co-legislation is slow; biotech competitiveness framing may water down Ch. VIII. Adoption ≠ immediate apply — phased enforcement typical. CLTR assessment: draft “delivers decisively” on screening vs their recommendations.

Evidence:

Analogue: EU AI Act — proposal 2021, adoption 2024, phased apply 2025–2027.

Would update if: Ch. VIII stripped in trilogue → revise adopted to <0.15; early adoption in 2026 → revise phased apply up.

Conf: M

P = 0.15 (2027-12) / 0.25 (2028-06) — Mandatory third-party CBRN eval (federal, blocking release)

Claim: Low probability of binding federal pre-release CBRN eval with block authority.

Why: Amodei Jun 2026 proposes FAA-style third-party testing at ~10²⁵ FLOP for bio/cyber/LoC — industry proposal, not law. EO 14409 (Jun 2 2026) explicitly prohibits construing EO as authorizing mandatory licensing/preclearance. Banks letter (Jun 2026) pushes voluntary RSI testing + CAISI visibility — adjacent, not bio-blocking. Bipartisan appetite for screening ≠ appetite for release blocks.

Evidence:

Analogue: FAA certification — exists for aircraft, not replicated for AI despite analogy.

Would update if: Mandatory CBRN eval with block authority passed and enforced ≥2 releases → falsifier in Node 2; revise all CBRN-mandate P upward.

Conf: M

P = 0.70 (2027-12) / 0.75 (2028-06) — State CBRN transparency + third-party assessment (RAISE + SB 53)

Claim: High probability of compliance prep by end-2027 and enforcement by mid-2028 for NY RAISE + CA SB 53 catastrophic-risk frameworks (incl. bioweapon assistance).

Why: RAISE signed Dec 2025, effective Jan 1 2027 — catastrophic risk includes expert-level bioweapon assistance; 72h incident reporting. SB 53 Encode-backed, Anthropic-endorsed. Law already enacted; P reflects prep + enforcement not passage. Tail: GAAIA federal preemption P≈0.25–0.35 by 2028 could weaken.

Evidence:

Analogue: CA privacy (CCPA) — state leads, federal lag.

Would update if: Federal preemption voids RAISE/SB 53 with no replacement → state path dead (Node 2 falsifier).

Conf: H

P = 0.30 (2027-12) / 0.45 (2028-06) — OSTP 50bp + function-based SOCs

Claim: Moderate-low probability OSTP framework hits 50bp + function-based sequences-of-concern on original timeline.

Why: Oct 2026 50bp deadline paused by EO 14292 (May 2025) per Node 2. OSTP procurement rule survived Biden EO partial revocation but implementation limbo. BMIA/EU may supersede OSTP rather than revive it — dual path increases cumulative P by 2028-06 to 0.45 without OSTP alone succeeding.

Evidence:

Analogue: OSHA standards often superseded by industry-consensus norms before rule finalizes.

Would update if: OSTP 50bp reinstated by executive action in 2026 → revise 2027-12 to ≥0.50.

Conf: L–M

P = 0.50 (2027-12) / 0.65 (2028-06) — IGSC v4 / function-based screening at scale

Claim: ~50% major IGSC protocol upgrade (v4-class) with function-based SOC pilots at scale by end-2027; 65% by mid-2028.

Why: IGSC v3.0 (2024) already baseline; NIST function-based R&D + IBBIS Common Mechanism commec pilots push v4. Jun 2026 letter from synthesis CEOs signals industry willingness ahead of law. Voluntary upgrade faster than statute — but “at scale” requires non-member adoption lag.

Evidence:

Analogue: PCI-DSS — industry standard ahead of federal mandate.

Would update if: BMIA mandates function-based SOC list → IGSC upgrade essentially certain; revise to ≥0.80.

Conf: M

P = 0.05 (2027-12) / 0.08 (2028-06) — Federal frontier pause (bio-motivated)

Claim: Near-zero federal pause driven specifically by bio/CBRN salience.

Why: Pause politically toxic (research_ai_pause_advocacy_playbook: FLI letter zero effect; natsec hawks anti-pause; Trump EO 14409 pro-innovation). Bio-motivated pause narrower than general pause — still loses to “beat China” + open-science framing. 0.05–0.08 allows black-swan (Tier-3 attack + election shock), not modal.

Evidence:

  • (internal note) §Executive summary, §1.4, §1.5
  • (internal note) §Modal path
  • EO 14409 (no mandatory licensing)

Analogue: 2023 pause letter — maximum awareness, zero law.

Would update if: Bipartisan bio-motivated training moratorium introduced with committee passage → revise to ≥0.15.

Conf: H

P = 0.10 (2027-12) / 0.15 (2028-06) — Biology FM weight gating (Evo-class)

Claim: Low probability of binding US rule restricting Evo-class open weights.

Why: Evo2 full open (Nature Mar 2026); Arc open-science default; AlphaFold3 gating debate showed pushback on weight restriction. No binding US biology-FM release rule exists (10_biosecurity_evo_dual_use.md §4). Policy momentum on (C) synthesis screening, not (B) FM gating — explicit Node 2 dual pathway.

Evidence:

Analogue: AlphaFold3 — gated weights, not banned; Evo took opposite default.

Would update if: Tier-3 attack traced to open Evo2 weights + bipartisan release-restriction bill → revise to ≥0.30.

Conf: L

P = 0.25–0.35 — GAAIA federal preemption of state AI laws (by 2028)

Claim: Tail risk that GAAIA-class bill preempts RAISE/SB 53 enforcement by 2028.

Why: GAAIA discussion draft (Jun 2026): frontier audits + 3-yr state development preemption. Cruz moratorium stripped 99–1 (Jul 2025) but preemption fight repeats (research_ai_pause_advocacy_playbook). ~$8.5M Q1 2026 lobbying on preemption. Not modal — states currently winning — but non-trivial tail.

Evidence:

  • (internal note) §1.2 (GAAIA, preemption)
  • (internal note) §RAISE tail

Analogue: Cruz BBB moratorium attempt — failed but revealed industry priority.

Would update if: Preemption passes → RAISE enforcement P drops sharply; Node 2 falsifier on state path.

Conf: M

Tail branches (conditional on Tier 2 live)

P = 0.55 — MODAL branch

Claim: ~55% of Tier-2 futures: BMIA/EU screening passes; RAISE/SB 53 reporting; IGSC upgrade; CBRN evals stay voluntary.

Why: Largest mass because screening coalition is broadest (synthesis industry + natsec + frontier CEOs + bipartisan Senate). Decoupling of physical-layer vs in silico governance is structural (timeline prediction node2 cbrn full §Modal synthesis). Assigned 0.55 as modal slice of {0.55, 0.20, 0.15, 0.10} — not independent draw from literature.

Evidence:

Analogue: Post-2001 airport security — mandatory physical screening, not “stop flying.”

Would update if: BMIA + EU both fail by 2028 despite Jun 2026 coalition → MODAL branch was wrong (Node 2 falsifier).

Conf: L (exact branch P); M (directional)

P = 0.20 — TAIL-A: Tier 3 low-skill before reg catches up

Claim: ~20%: first documented low-skill attack or classified end-to-end demo before screening deployed at scale.

Why: Evo2 ASR scaling; DBTL closure; non-IGSC vendors; 50bp delay create window. Not modal because RAND + no public Tier-3 yet, but GeneBreaker shows skilled-actor path is live — low-skill is extrapolation not confirmation. 0.20 = substantial tail, not majority.

Evidence:

Analogue: Aum Shinrikyo — small group, high skill; Tier-3 tail asks if skill floor drops further.

Would update if: Documented Tier-3 attack → reweight TAIL-A up, MODAL down.

Conf: L

P = 0.15 — TAIL-B: Regulatory delay (12–18 mo after near-miss)

Claim: ~15%: near-miss or incident occurs but rules bite only after 12–18 mo lag.

Why: Trump deregulation; OSTP limbo; preemption fights; industry preference for voluntary compliance. COVID policy lag analogue in Node 2 hybrid rule. Smaller than MODAL because Jun 2026 coalition may front-run incident.

Evidence:

  • EO 14409
  • (internal note) §1.2
  • (internal note) §TAIL-B, §Hybrid rule

Analogue: COVID test rollout — capability existed, regulatory response lagged.

Would update if: BMIA signed within 6 mo of near-miss → TAIL-B overestimated.

Conf: L–M

Would update if: US–CN joint synthesis treaty (P≈0.03) somehow passes → TAIL-C collapses.

Conf: L

Actor table

P = 0.70 — US executive: voluntary 30-day cyber/frontier review (EO 14409)

Claim: ~70% that covered frontier developers participate in voluntary pre-release review framework through 2027.

Why: EO 14409 (Jun 2 2026) creates structured voluntary pathway; soft pressure from contracts, “trusted partner” access, natsec engagement. Not 0.90+ because participation explicitly non-mandatory and some labs may opt out (Meta, open-weight shops).

Evidence:

Analogue: Voluntary SOC 2 — market pressure drives compliance without mandate.

Would update if: Major lab publicly refuses all EO 14409 engagement with no commercial penalty → revise to ≤0.50.

Conf: M

P = 0.12 — US executive: mandatory CBRN licensing blocked / prevented

Claim: ~12% that executive branch actively blocks or prevents mandatory CBRN licensing if Congress moves — low but non-zero veto/agency-nullification tail.

Why: EO 14409 already constrains EO-based mandate. Trump admin would likely veto bio-blocking bill if rare passage occurs. 0.12 = executive successfully kills mandate, not P(mandatory CBRN passes).

Evidence:

  • EO 14409 § “Nothing… mandatory licensing”
  • (internal note) §Actor table (US executive)
  • (internal note) §1.2

Analogue: Sustained veto on natsec tech — rare.

Would update if: CBRN block bill passes both chambers with veto-proof margin → estimate wrong.

Conf: L–M

P = 0.35 — US executive: OSTP 50bp revision proceeds

Claim: ~35% executive branch revives or revises OSTP 50bp + function-based SOC timeline by end-2027.

Why: EO 14292 paused Oct 2026 deadline; BMIA may supersede OSTP track. Commerce-led BMIA fits admin preference.

Evidence:

Analogue: Deferred agency rules revived under new lead agency.

Would update if: BMIA signed → OSTP revision moot.

Conf: L–M

P = 0.45 — US Congress: bio screening legislation (BMIA-class)

Claim: ~45% BMIA-class mandatory screening enacted by end-2027 — same as modal regulatory P, Congress actor slice.

Evidence:

Analogue: See BMIA modal section above.

Would update if: Committee markup in 2026 → ≥0.55.

Conf: M

P = 0.20 — US Congress: explicit CBRN eval mandate

Claim: ~20% Congress passes explicit federal CBRN third-party eval with enforcement by 2028.

Why: Industry unites against release blocks; screening coalition does not extend. Amodei proposal not yet bill.

Evidence:

Analogue: SB 1047 veto path.

Would update if: Bipartisan CBRN eval bill with lab support → revise up.

Conf: M

P = 0.05 — US Congress: frontier pause

Claim: ~5% Congress enacts frontier pause by 2028.

Evidence:

  • (internal note) §1.2, §1.4
  • (internal note) §Actor table

Conf: H

P = 0.50 — DNA synthesis industry: major IGSC upgrade (v4-class)

Claim: ~50% IGSC delivers major protocol upgrade by end-2027.

Evidence:

Conf: M

P = 0.30 — DNA synthesis industry: compliance without law

Claim: ~30% enhanced screening covers ≥80% US synthesis volume before federal mandate.

Why: IGSC members + letter signatories; non-members and benchtop gap cap at 0.30.

Evidence:

  • (internal note) §2.3
  • (internal note) §Actor table

Conf: L–M

P = 0.75 — Frontier labs: CBRN mitigations ship

Claim: ~75% frontier labs ship visible CBRN mitigations on major releases through 2027.

Why: RSP / Preparedness / Sleeper Agents → patch not pause.

Evidence:

  • (internal note) §2.1
  • (internal note)
  • (internal note) §Actor table

Conf: M–H

P = 0.02 — Frontier labs: stop training

Claim: ~2% major frontier lab stops training (bio-motivated) by 2028.

Evidence:

  • (internal note) §1.3
  • (internal note) §Actor table

Conf: H

P = 0.40 — Frontier labs: public Tier-2 disclosure under RAISE/SB 53

Claim: ~40% public disclosure of Tier-2-class finding via reporting/leak by 2028.

Why: RAISE reports confidential to NY DFS; public = leak + strategic transparency subset.

Evidence:

Conf: L–M

P = 0.60 — China: domestic screening enforcement ↑

Claim: ~60% CN tightens synthesis screening enforcement through 2028.

Evidence:

  • (internal note) §Actor table (China)
  • (internal note) §1.4

Conf: L–M

P = 0.08 — China: FM weight gating

Claim: ~8% CN binding restrictions on open biology FM weights.

Evidence:

  • (internal note) §4
  • (internal note) §Actor table

Conf: L

P = 0.03 — China: US–CN joint bio-AI treaty

Claim: ~3% formal joint treaty by 2028.

Evidence:

  • (internal note) §1.2, §1.4

Conf: L

P = 0.55 — EU: Biotech Act adoption with screening intact (by 2028)

Claim: ~55% cumulative by 2028 — aligns actor table with modal EU track extended horizon.

Evidence:

Conf: M

P = 0.30 — EU: CBRN eval mandate for GPAI

Claim: ~30% EU mandates CBRN eval for GPAI/high-risk in biotech context by 2028.

Evidence:

  • (internal note) §3 (FAR.AI EU CBRN consortium)
  • (internal note) §Actor table

Conf: L–M

P = 0.08 — Public: mass bio-x-risk protest

Claim: ~8% sustained mass protest framing AI bio-x-risk by 2028 (no Tier-3).

Evidence:

  • (internal note) §1.4
  • (internal note) §Actor table

Conf: L–M

P = 0.25 — Public: sustained media >2 mo given Trigger E

Claim: Conditional on Trigger E, ~25% media sustains >2 months.

Evidence:

  • (internal note) §Actor table, §Trigger E

Conf: L

p(doom) conditional claims

P(extinction | MODAL branch) — Low (~1–3% of all extinction mass by 2050)

Claim: MODAL governance leaves AI-enabled bio at ~1–3% of total extinction mass by 2050 (not misalignment path).

Why: Screening + RAISE shrink skilled-actor tail; Adelstein ~8–10% bio-inclusive total — MODAL discounts slice.

Evidence:

  • (internal note)
  • (internal note) §p(doom) link

Analogue: Nuclear terrorism — gated by access.

Would update if: Tier-3 under full MODAL rules → estimate failed.

Conf: L (quant); M (direction)

P(extinction | TAIL-A) — Moderate–high within bio bucket

Claim: Tier 3 before screening → moderate–high within Adelstein bio bucket; single pandemic tail, not misalignment.

Evidence:

  • (internal note) §TAIL-A, §p(doom)
  • RAND RBA4087-1

Conf: L

P(extinction | TAIL-B) — Moderate

Claim: 12–18 mo delay after near-miss → moderate bio extinction risk during gap window.

Evidence:

  • (internal note) §TAIL-B

Conf: L

P(extinction | TAIL-C) — Moderate

Claim: Open weights + arbitrage + eval gaming → moderate chronic risk.

Evidence:

  • (internal note)
  • (internal note)
  • (internal note) §TAIL-C

Conf: L

P(Tier 3 before adequate screening | Tier 2 by 2027) > ~0.15 — crux

Claim: Above ~15%, bio bucket dominates misalignment in Adelstein split (~2.6% misalignment vs ~5–8pp bio).

Why: TAIL-A weight 0.20 minus partial mitigation ≈ crux threshold. Weight by branch P; no double-count.

Evidence:

  • (internal note) §Crux
  • (internal note) §Why community numbers differ

Would update if: BMIA + GeneBreaker mitigations → P(Tier 3 | Tier 2) <0.10.

Conf: M (logic); L (threshold)

P(weighting) — No double-count MODAL + tails

Claim: Weight branch p(doom) claims by P(MODAL)=0.55, P(TAIL-A)=0.20, etc.

Evidence:

  • (internal note) §Double-counting guard

Conf: H

Sandbagging cross-cut

P(confound) — Public evals alone unreliable for Tier assignment

Claim: Tier 1–2 from public WMDP scores are lower bounds; sandbagging/eval-awareness inflates apparent safety.

Why: Goodfire +3–18 pp refusal when aware; elicited sandbagging −35–40% WMDP (evaluation_awareness_llms.md).

Evidence:

Conf: H

Index

CategorySections
Hybrid timing5
Tier status4
Trigger E6
Modal regulatory10
Tail branches4
Actor table18
p(doom)6
Sandbagging1
Total54

External sources (consolidated)

Phase 2b — Synthetic media / epistemic collapse & BMIA politics (crosscut §3)

Source: crosscut secondary cruxes §3 — ranked #7 leverage on ~17% doom region (emergent sim) (+1–3pp if BMIA-fail tail).
Parent link: Extends P(BMIA)=0.45 load-bearing P #2; MODAL branch 0.55; TAIL-B regulatory delay 0.15.

Claim — democracy can’t pass BMIA when epistemic infrastructure collapses

By 2027–2028, synthetic media + AI-generated persuasion degrades democratic epistemic infrastructure enough that physical-layer consensus (BMIA, screening mandates) fails in Congress — even when insider/natsec coalition favors it. Not extinction-direct; enables bio tail by blocking Cluster B (correlation matrix).

Node 2 bottleneck: BMIA bipartisan (Cotton–Klobuchar Jan 2026); screendna coalition — elite consensus exists (§P=0.45 BMIA modal); public epistemic environment is the bottleneck for floor vote, not committee drafting.

P(modal / tail) — epistemic outcomes and BMIA timing

OutcomePWindowBMIA effect
Modal: Epistemic noise slows BMIA ~6–12 mo; passes by 20280.552027–2028P(BMIA by 2027-12) stays ~0.45; slips to mid-2028 (§0.60 horizon)
Tail: Epistemic collapse — BMIA fails 119th Congress; screening stays patchwork0.302027-12Load-bearing P #2 ↓ to ~0.25–0.30
Tail: “Chernobyl for trust” — election/deepfake crisis → punctuation that helps screening coalition0.152026–2028P(BMIA) ↑ to ~0.55–0.65; strange-bedfellows

Interaction with load-bearing P #2:

P(BMIA by 2027-12) baseline = 0.45  (§Modal regulatory BMIA section)
This crux explains downside mass on that estimate — epistemic tail = primary BMIA-fail mechanism
P(BMIA-fail | epistemic collapse tail) ≈ 0.55–0.65 conditional

Correlation warning: Do not multiply P(epistemic tail) × P(BMIA fail) × P(no pause) — use scenario branches (my pdoom §Branch model).

Evidence — survey vs action, Kingdon, political deepfakes

StreamDetailBMIA relevance
Survey vs action gap70–84% support AI regulation; no federal pause after FLI 2023Broad support ≠ floor vote under noise — ai_governance_political_strategy.md §Executive summary
Kingdon problem streamWeak for x-risk; moderate for deepfakes/jobs/child safety — §1.4BMIA is narrower than pause but still needs trustworthy deliberation
Political deepfakes84% support ban (ai_governance_political_strategy.md); Grok non-consensual deepfake crisis Jan 2026Salience ↑, institutional trust ↓ — dual effect
Open Q §6.3Non-catastrophic focusing event (deepfake election, AI flash crash) — may punctuate or polarizeMaps to 0.15 punctuation tail vs 0.30 collapse tail
Node 4 media fatigueDowns 6–10 wk peaks (Node 4 §20) — synthetic flood accelerates fatigueBMIA debate drowned in synthetic noise
Elite vs public gapscreendna Jun 2026 — CEOs + natsec for screeningCongress needs legible public consensus — epistemic collapse breaks legibility

Epistemic collapse — delaying the BMIA coalition

Mechanism: BMIA coalition = synthesis industry + natsec + frontier CEOs + bipartisan Senate (elite). Passage requires:

  1. Committee markup without poison pills
  2. Floor time not consumed by higher-salience synthetic crises
  3. Credible technical debate — not dismissed as “deepfake transcript of Cotton”
  4. Governor/executive not vetoing under election-year chaos

Epistemic collapse tail (P=0.30) breaks (2) and (3):

Failure channelExampleEffect on P(BMIA by 2027-12)
Synthetic flood drowns committee debateFake “leaked” BMIA draft with absurd capsMarkup delayed → TAIL-B 0.15 overlap
Disinformation targets sponsorsFake Klobuchar/Cotton quotes on “AI ban”Bipartisan bridge fractures
Public can’t distinguish real vs synthetic expert testimonyDeepfake Amodei “oppose screening”Industry endorsement neutralized
Media fatigueNode 4 §20 — 6–10 wk peaksBMIA becomes “another AI panic”

Link TAIL-B (§P=0.15): Near-miss or insider salience occurs but rules bite only after 12–18 mo lag — epistemic collapse extends TAIL-B by blocking focused post-near-miss window.

Link MODAL branch (§P=0.55): BMIA passes but 6–12 mo late — Tier 2 live window (§Hybrid timing) unchanged; P(Tier 3 before screening) load-bearing P #3 ↑ modestly.

Deepfake election cycle — 2026–2028 punctuation risk

Claim: 2026–2028 US election cycles intersect peak synthetic-media capability — bifurcated tail:

PathPMechanismBMIA
Collapsesubset of 0.30 tailElection-year synthetic chaos → no trustworthy legislative processBMIA dies in committee
Punctuation (“Chernobyl for trust”)0.15Deepfake election crisis → creatives + natsec + synthesis industry strange bedfellowsP(BMIA) ↑ — ai_governance_political_strategy.md §5 labor + deepfake + natsec package

Evidence — Grok Jan 2026: Non-consensual deepfake crisis — salience ↑, institutional trust ↓ simultaneously. Net direction uncertain — explains both tails in same event class.

Deepfake election scenarios:

2026 midterms / 2028 presidential cycle:
  - Synthetic candidate audio/video → litigation, not legislation
  - "Everything is fake" cynicism → voters discount **real** BMIA benefits
  - OR: High-profile election deepfake → NO FAKES / ELVIS coalition **bridges** to BMIA floor vote

Node 6 gridlock interaction: Epistemic collapse reinforces Node 6 §Gridlock 0.45 — federal dysfunction on all AI bills, not just GAAIA; BMIA physical-layer status may not exempt it from epistemic veto (unlike natsec classified track).

Democracy capacity crux — load-bearing P #2 decomposition

Crux statement: US democracy can pass narrow physical-layer bio screening iff epistemic infrastructure supports technical policy debate. Synthetic media attacks process legitimacy, not screening science.

Load-bearing PBaselineIf epistemic collapse tail (0.30)If punctuation tail (0.15)
P #2 P(BMIA by 2027-12)0.45~0.25–0.30~0.55–0.65
P #3 P(Tier 3 before screening | Tier 2)crux >0.15↑↓
Bio tail branch 0.17×35%priced↑ mass on bio tail↓

Not extinction-direct: Epistemic collapse does not cause misalignment; enables bio tail by blocking Cluster B (physical mitigation).

p(doom) direction:

PathEffect on doom region (emergent sim)
BMIA-fail epistemic tail+2–4pp — bio tail branch 0.17×35% more likely
Punctuation tail−1–2pp — BMIA coalition punctuated
Modal delay only+0–1pp — absorbed in existing 2028-06 P=0.60

Would update if — falsifiers and confirmers

ObservationDirection
BMIA dies in committee with documented synthetic-media disinformation campaign against sponsors; no substitute by 2028Up (collapse tail)
BMIA signed despite election-year synthetic floodDown (collapse) — modal delay path confirmed
EU Biotech Act spillover forces US patchwork harmonizationDown (collapse) — external punctuation bypasses US epistemic failure
Major authenticated-media infrastructure (C2PA-scale adoption) restores baseline trust for technical policy debateDown (collapse)
Deepfake election crisis → NO FAKES + BMIA joint floor package passesUp (punctuation tail)
Trigger E near-miss (§P=0.25–0.35) overcomes epistemic noiseDown (collapse) — bio salience punctuates through synthetic fog
NodeLink
N2 §P=0.45 BMIALoad-bearing P #2 — this crux explains downside mass
N2 §Trigger E near-missP=0.25–0.35 — may puncture epistemic fog (punctuation path)
N2 §TAIL-BRegulatory delay 0.15 — epistemic collapse extends
N4 §20 MediaFatigue; x-risk competes with jobs/deepfakes
N6 §Gridlock0.45 — epistemic collapse reinforces federal gridlock
N1Labor shock coalitions — potential bridge (labor + deepfake) per ai_governance_political_strategy.md §5
correlation matrixCluster B physical mitigation vs Cluster C BMIA delay

Phase 2b status (Node 2): 8 sections appended 2026-07-04. Crosscut §3 — zero omissions.