← Evidence index · 中文 · Main post
Each section: Claim · Why · Evidence · Analogue · Would update if · Conf (H/M/L).
Parent: Node 6 policy brief
P = 0.12 — GAAIA passes with 3-year state preemption intact
Claim: House/Senate passes GAAIA-class bill including 3-year preemption of state “model development” rules.
Why: Strongest federal safety text yet (Jun 2026 draft) but preemption fought same as Cruz moratorium — industry wants one federal rule; states resist. 12% = tail not modal: Encode + unions + Newsom-style governors historically block pure preemption.
Evidence:
- (internal note) — GAAIA Jun 4 2026; Cruz stripped 99–1 Jul 2025
- (internal note) §T2 P=0.12 (same event)
- (internal note) §3.4 GAAIA vs SB 53 table
- ~$8.5M Q1 2026 lobbying on preemption (playbook)
Analogue: ACA vs state health laws — partial preemption fights took years.
Would update if: Bipartisan markup keeps preemption + White House endorsement.
Conf: M
P = 0.18 — GAAIA passes, preemption stripped
Claim: Federal audit/reporting law passes without gutting SB 53/RAISE dev obligations.
Why: Compromise path — Banks/Trahan safety text + Wiener/Encode strip preemption (mirrors SB 1047→SB 53 negotiation). More likely than full preemption win.
Evidence:
- SB 1047 veto → SB 53 signed (transparency survives, caps die)
- Anthropic endorsed SB 53, criticized heavy-handed caps
- node4 GAAIA passes House 0.25 — subset is stripped-preemption pass
Conf: M
P = 0.45 — Gridlock; states primary
Claim: No federal law by 2028; SB 53 + RAISE enforce as designed.
Why: Default US federal dysfunction on tech; Trump EO voluntary framework may satisfy industry enough to kill GAAIA urgency. States already moving (SB 53 law, RAISE effective 2027-01-01).
Evidence:
- FLI 2023 → zero federal law
- EO 14409 Jun 2026 as federal substitute
- (internal note) RAISE P=0.70 prep
Conf: M
P = 0.25 — GAAIA dies; EO-only federal framework
Claim: GAAIA fails; Trump-class voluntary review becomes de facto federal standard.
Why: Industry preference for voluntary + export controls over statutory audits. OpenAI alignment with lighter EO.
Evidence:
- Playbook §Trump EO 2026-06 — explicitly not mandatory licensing
- Same-month cluster: EO + GAAIA + Banks letter — EO may satisfy moderate Rs
Conf: M
P = 0.55 — Unions + populist right block federal preemption (conditional on GAAIA floor)
Claim: If GAAIA reaches floor, ~55% chance preemption stripped or bill dies (joint with Encode).
Why: Jul 2025 Cruz moratorium stripped 99–1 — left + right united against federal ban on state AI action. Preemption is different frame but overlapping coalition.
Evidence:
- node1 anti-preemption 0.55
- Playbook §1.4 coalitions
Conf: M
p(doom) — P(no effective pause) +1–3pp if O1
Claim: Federal preemption win increases P(coordination failure on pause) modestly.
Why: Removes state-level “laboratory” for harder rules; centralizes to DC where pause is dead. Does not directly cause extinction.
Evidence:
- Cross-node correlation matrix — GAAIA ↔ P(no pause)
Conf: L–M
Phase 2b — AI-lab-as-state / classified defense stack (crosscut §4)
Source: crosscut secondary cruxes §4 — ranked #4 leverage on ~17% doom region (emergent sim) (+2–4pp if natsec capture tail).
Parent link: Extends Node 6 GAAIA/preemption; ties to Node 4 §10 E4 (P=0.37), Trigger E2 classification fight.
Claim — alignment migrates to classified/natsec channel
Frontier AI governance migrates to classified/natsec channel: Palantir/Anduril-style integration, Pentagon deals (xAI Feb 2026), CAISI/Banks visibility demands, NSA vs CAISI turf. Alignment debates happen in secret — increasing P(E4), decreasing public Trigger E, increasing race + modal oversight without halt.
Node 6 framing: GAAIA/EO modal path (§Gridlock 0.45, §EO-only 0.25) assumes public federal rulemaking. Classified stack crux says parallel track may absorb C10 eval fights — weakening public preemption fight while hardening E4 no-leak path.
P(modal / tail) — dual track vs natsec capture vs transparency win
| Outcome | P | Notes | Node 6 / N4 mapping |
|---|---|---|---|
| Modal: Dual track — public voluntary EO + classified evals | 0.60 | Jun 2026 stack (EO 14409, GAAIA, Banks, OpenAI blueprint) | Reinforces §Gridlock 0.45 + EO-only 0.25 combined public track |
| Tail: Natsec capture — alignment classified; public gets sanitized RSI essays only | 0.28 | E4 path amplified | ↑ Node 4 §10 P(E4) from 0.37 — not independent; conditional boost |
| Tail: Transparency win — CAISI public benchmarks + anti-NSA-solo (Banks framing) holds | 0.12 | Encode/state allies | ↓ E4; ↑ Trigger E1/E2 if public benchmarks leak concern |
Conditional dynamics:
P(public Trigger E | natsec capture tail) << P(public Trigger E | modal dual track)
P(E4 | natsec capture) ≈ 0.50–0.55 (vs baseline 0.37)
P(GAAIA with natsec exempt deployment category) ↑ if capture tailEvidence — xAI, Banks, OpenAI blueprint, AISI→CAISI
| Stream | Detail | p(doom) direction |
|---|---|---|
| xAI Pentagon Grok (Feb 2026) | “All lawful purposes” — no Anthropic-style safeguards | ↑ race; ↓ public alignment leverage |
| Banks letter (Jun 2026) | RSI testing, unpublished internal models visibility, CAISI vs NSA | Natsec hawk, not pause — visibility not veto |
| OpenAI blueprint (Jun 2026) | CAISI mandatory eval, RSI indicators — visibility not veto; preemption negotiation | Supports §P=0.18 stripped GAAIA; not halt |
| AISI → CAISI rebrand (Jun 2025) | Shift from broad societal risk to innovation/natsec | Modal dual-track evidence |
| Node 4 §10 E4 | Natsec classification ↑ criminal exposure for leakers; managed disclosure cheaper | P(E4)=0.37 baseline — capture tail pushes higher |
| Node 3 | Theft crisis → more classified evals, same P(halt) | Race cluster hot without brake |
Sources: (internal note); research_RSI_june2026_glossary_and_responses.md; research_recursive_self_improvement_SOTA_2026.md §4.
E4 no-public-leak path — amplified under classified stack
Mechanism: When C10 evals move SCI/SAP-adjacent, E4 becomes cheaper than E1/E2:
| Factor | Effect on E4 | Evidence |
|---|---|---|
| Criminal exposure for leakers | ↑ E4 | CFAA/espionage risk on eval dump (Node 4 §8 E2) |
| Managed disclosure substitute | ↑ E4 | Anthropic RSI (2026-06) — corporate-framed warning same month as EO/GAAIA |
| Whistleblower pipeline thinning | ↑ E4 | Sharma “exit not organize” (2026) — Node 4 §10 |
| Banks visibility without public release | ↑ E4 | Regulators see evals; public sees sanitized summary |
| No training halt from classified review | Modal | Classified eval ≠ pause — parallel to Saunders → transparency not compute |
Quantitative sketch: Baseline P(E4)=0.37 (Node 4 §10). Under natsec capture tail (P=0.28 of this crux), conditional P(E4 | capture) ≈ 0.50–0.55 — public Trigger E mass shifts to E4.
Policy consequence for Node 6: GAAIA public audit text may pass (§P=0.18 stripped) while C10 scheming evals sit in classified annex — modal oversight without halt (Cluster A, correlation matrix).
Palantir / Anduril — classified stack and revolving door
Institutional fusion (not necessarily ideological capture):
| Entity | Role | Timeline relevance |
|---|---|---|
| Palantir | Data fusion, natsec analytics, classified contracts | Template for lab-as-state integration — eval visibility to IC without public leak |
| Anduril | Defense AI, autonomous systems, Lattice | Defense-AI stack vendor class — competes with frontier labs for classified deployment slots |
| MTS pipeline | Anduril → Anthropic Applied AI (reported) | Talent revolving door — institutional fusion |
| Alex Bores | ex-Palantir, RAISE (NY) | State-level vetting bridge — public track; different from classified eval |
| Kevin Zhu / Doom Debates | CS students avoid Palantir on liberal bias | Polarized talent/politics — AI safety coded TESCREAL-left → right-flank anti-pause alliance space |
Natsec capture tail (P=0.28): Alignment/scheming evals default SCI/SAP-classified; public gets RSI essays; Palantir/Anduril-class integrators run classified eval pipeline — not frontier lab safety team with public veto.
Counter-tail transparency win (P=0.12): Banks framing blocks NSA-only lane; CAISI publishes redacted scheming eval summaries — forces Node 4 Trigger E2 despite classification fight.
Defense-AI stack — alignment debates in secret
Claim: Under natsec capture, alignment debates happen in secret — decreasing public Trigger E, increasing race + modal oversight without halt.
| Dimension | Public track | Classified / defense-AI track |
|---|---|---|
| Scheming eval results | Sanitized blog (RSI essay class) | Full probes in SAP compartment |
| Halt decision | Conditional pause rhetoric (no verify) | No public halt; IC continuity mission |
| Whistleblower Trigger E | E1 memo, E2 dump | Suppressed — E4 modal |
| GAAIA / EO | Voluntary review, audit reporting | Exempt natsec deployment category (tail risk) |
| Race narrative | ”Beat China” public | Classified eval accelerates mil deployment without public scare |
Alignment in secret — p(doom): Public cannot apply political pressure for halt what they cannot see; ↓ P(organized pause); ↑ P(no shutdown at ASI) load-bearing P #6. Whistleblower E2 (eval dump) despite classification would shift back to Node 4 tail — low P under capture.
Interaction Node 2: Natsec framing helps BMIA (physical layer) while hurting CBRN eval mandates — split governance (crosscut secondary cruxes §4 links table).
p(doom) direction — natsec capture vs transparency win
| Path | Effect on doom region (emergent sim) | Cluster |
|---|---|---|
| Natsec capture tail (0.28) | +2–4pp | Cluster A — P(no pause), P(E4), P(race post-theft) ↑; public whistleblower ↓ |
| Modal dual track (0.60) | +0–1pp | Public EO + classified evals — already partially in Node 6 base rates |
| Transparency win tail (0.12) | −1pp | CAISI public benchmarks; anti-NSA-solo holds |
Rough: Natsec capture vs transparency win → +2–4pp vs −1pp on extinction (crosscut secondary cruxes §4).
Node 6 O1 preemption interaction: Capture tail increases industry preference for single federal classified framework (§P=0.12 full preemption or EO-only 0.25) — state vetting labs (SB 53) bypassed for natsec deployment category.
Would update if — falsifiers and confirmers
| Observation | Direction |
|---|---|
| C10 evals SCI/SAP-classified by default | Up (capture) |
| GAAIA passes with exempt natsec deployment category | Up (capture) |
| Second frontier lab signs unrestricted classified contract (xAI-class) | Up (capture) |
| CAISI publishes redacted scheming eval summaries with statutory basis | Down (transparency) |
| Banks/OpenAI blueprint blocks NSA-only testing lane | Down (transparency) |
| Whistleblower E2 (eval dump) despite classification | Down (capture) → Node 4 tail |
Links to existing nodes
| Node | Link |
|---|---|
| N4 §10 E4 | P(E4)=0.37; classification ↑ leak cost |
| N4 §8 E2 | Eval dump — classification fight |
| N6 | GAAIA O1 preemption; CAISI codification; Trump EO modal |
| N3 | Export controls + classified evals post-theft |
| N2 | Natsec helps BMIA physical layer; hurts CBRN eval mandates |
| correlation matrix | Cluster A acceleration / race |
Phase 2b status (Node 6): 8 sections appended 2026-07-04. Crosscut §4 — zero omissions.