Long Yi
← Back to all writing

p(doom) evidence — Node 11: corporate governance hollowing

July 5, 2026

Predictionp(doom)evidencenode-11

← Evidence index · 中文 · Main post

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

Parent: node11 corporate governance
Purpose: Claim | Why | Evidence | Analogue | Would update if | Conf — for every Node 11 probability.

1. Node definition — governance structure is load-bearing for halt

Claim: Frontier lab corporate governance (board independence, binding PBC/LTBT provisions, safety reporting lines) is a separate crux from whistleblower events (Node 4): it sets P(halt | internal crisis) ≈ 0.12, not ~0.40 naive prior.

Why: Node 4 models human response to leaks; Node 11 models institutional default leaks face. Board crisis 2023 + Leike 2024 + Sharma 2026 = repeated observation of zero training stop despite internal leverage.

Evidence:

  • (internal note) §29–31 — board crisis, Leike, Saunders
  • (internal note) §2 — hollowing branch 0.65
  • (internal note) §2.3, §5 — crisis timeline

Analogue: 2008 financial crisis — board risk committees existed on paper; no trading halt when models broke.

Would update if: ≥2 labs halt >90d after pure governance crisis (no autonomous harm).

Conf: H

2. Branch M1 — modal hollowing P = 0.65

Claim: P(continued safety-team hollowing + narrative without veto, 2026–28) = 0.65.

Why: Default track 2024–26: superalignment dissolved, safety absorbed into product divisions, conditional pause rhetoric without halt, Sharma exit not organize. Oct 2025 OpenAI recap and Apr 2026 LTBT majority are paper upgrades atop practice flat.

Evidence:

  • (internal note) §23 OpenAI — “actions contradicted stated commitments”
  • (internal note) §23 — P(public halt >30d) = 0.08 modal
  • OpenAI post-board-crisis: training continued (node4 §29)

Analogue: Pharma — pharmacovigilance teams exist; blockbuster launch pressure overrides rare drug withdrawals.

Would update if: Documented ≥2 labs restore independent safety org reporting to CEO-only with deploy veto used once.

Conf: M

3. Branch T1 — board crisis repeat without halt P = 0.25

Claim: P(acute OpenAI-2023-class governance fight without training stop by 2028) = 0.25.

Why: Single strongest analogue: board had firing power, 5-day media storm, safety-aligned directors lost — training never paused. Base rate for second crisis at OpenAI or peer given capex/Stargate pressure: ~1 in 4 over 2 yr window.

Evidence:

  • (internal note) §2.3 — Altman fired 2023-11-17, reinstated 2023-11-21, Toner out
  • (internal note) §29 — “insiders with leverage still lost”
  • NBC News — OpenAI restructuring Oct 2025 — commercial/IPO pressure continues

Analogue: WeWork 2019 — board challenge → CEO returns → governance reverted.

Would update if: OpenAI Foundation removes Altman and new CEO announces 30d training pause.

Conf: M

4. Branch T2 — safety re-empowered P = 0.10

Claim: P(safety team re-empowered — hard RSP/FSF stop invoked + visible halt or deploy block by 2028) = 0.10.

Why: Counter-tail reserved for observed invocation, not rhetoric. Anthropic RSP v3.0, OpenAI SSC, DeepMind FSF v3.1 create mechanisms; zero public invocations 2024–26. LTBT majority (Apr 2026) theoretically enables block — Sharma exit same quarter undercuts insider confidence.

Evidence:

  • Anthropic RSP v3.0 (Feb 2026) — self-enforced, no external accountability (research_ai_governance_landscape.md)
  • Anthropic LTBT announcement
  • Node 4 §23 — P(RSP hard stop) = 0.12 modal, 0.25 tail-gov

Analogue: Nuclear SCRAM — designed to work; rarely tested in production crisis.

Would update if: Public RSP ASL-4+ trigger → documented training slowdown + Risk Report.

Conf: L–M

5. P(halt >30d | internal governance crisis) = 0.12

Claim: P(any frontier lab public training halt >30 days | internal governance crisis without autonomous harm) = 0.12 (range 0.08–0.18).

Why: Product of conditional chain in node11 corporate governance — board favors safety ~15% × eval trigger ~35% × low race-pressure counterfactual ~20%. Four historical crises → zero halts.

Evidence:

  • Node 4 §29–31 — board, Leike, Saunders
  • (internal note) §5.2 — Sharma 2026 no halt
  • Crosscut §2 — derived P(halt|crisis) 0.10–0.15

Analogue: node4 composite prior on lab halt | scandal.

Would update if: Leike-class resignation immediately followed by named run pause (never observed).

Conf: M

6. P(≥2 labs halt >30d | governance crisis) = 0.03

Claim: P(multi-lab coordinated halt >30d | governance crisis) = 0.03.

Why: Node 4 §23 — P(≥2 labs halt) = 0.18 × 0.15 tail-gov ≈ 0.03 in whistleblower frame; governance-only crisis weaker trigger than C10 eval — same order of magnitude, not higher.

Evidence:

  • (internal note) §23, §308–309
  • Anthropic RSI — multilateral verify prerequisite (Anthropic_When_AI_Builds_Itself_2026.md §1)
  • Seoul 2024 Frontier Commitments — voluntary, no halt

Analogue: COVID vaccine pause (AZ thrombosis) — single product hold, not industry halt.

Would update if: Seoul-style summit produces binding synchronized halt clause with verification.

Conf: M

7. OpenAI board crisis 2023 — training continued (anchor)

Claim: OpenAI Nov 2023 crisis is the strongest downward bound on P(halt | scandal): board power + insider coalition → commercial default.

Why: Even structural governance failure (CEO fired) did not pause GPT-4-era training. Sets prior for all subsequent lab crises.

Evidence:

  • (internal note) §2.3–2.4 — “安全派输了董事会战争”
  • (internal note) §29
  • (internal note) — board crisis → commercial default

Analogue: Stronger than FLI 2023 (no insider leverage) for modal calibration.

Would update if: FOIA/SEC filing shows undisclosed training pause during Nov 17–21 window (no public evidence).

Conf: H

8. Leike / superalignment dissolution — zero operational leverage

Claim: Jan Leike May 2024 resignation + superalignment team dissolution moved narrative but zero operational leverage on OpenAI training.

Why: Viral tweet (“safety culture lost to shiny products”); 20% compute pledge abandoned; team absorbed. Cheaper costly signal than Saunders (no memo dump) — between E4 and E1 in Node 4 taxonomy.

Evidence:

  • (internal note) §3.2 — Leike 2024-05-17, superalignment dissolved
  • (internal note) §23 OpenAI — compute pledge denied
  • Node 4 §31 — “Insider moral authority ≠ operational leverage”
  • TechCrunch — superalignment withered

Analogue: Executive resigning on principle — stock price blip, no strategy change.

Would update if: OpenAI reconstitutes independent superalignment-class team with published compute budget.

Conf: H

9. Sharma exit 2026 — exit not organize; pipeline thins

Claim: Mrinank Sharma Feb 2026 exit (Sleeper Agents co-author → leave industry for poetry) increases P(thin whistleblower pipeline) and decreases P(costly internal leak).

Why: Breaks “migrate to Anthropic” pattern (Leike, Schulman). Signals structural despair vs company-specific. Safeguards Research lead vacancy without public succession narrative.

Evidence:

  • (internal note) §5.1–5.2 — Sharma 2026-02-09
  • (internal note) — Sleeper Agents, Constitutional Classifiers lineage
  • Node 4 §24 — “Sharma 2026: exit not organize — defection from insider advocacy”
  • Node 4 §10 — E4 ↑ as managed disclosure substitute

Analogue: Climate scientist leaving academia for farming — field loses advocate, not just employer loses employee.

Would update if: Sharma or peer publishes documented leak with policy effect (Kokotajlo-class).

Conf: M

10. Anthropic RSI essay — E4 managed disclosure; continue training

Claim: Anthropic Institute When AI Builds Itself (2026-06-04) is primary E4 path: internal RSI data + conditional expect pause — same month Trump EO + GAAIA acceleration frame wins.

Why: Wording we expect that we would not commit; verification mechanism does not exist; publication preempts whistleblower pressure while training implied continue. Node 4 P(E4)=0.37 anchor.

Evidence:

  • (internal note) — §1 conditions, §5.2 tension (8× code velocity + expect pause)
  • (internal note) §5.1 — 2026-06-04 entry
  • Node 4 §32 — “Labs preempt whistleblower with managed disclosure”

Analogue: Tobacco industry “we support regulation” while opposing binding rules.

Would update if: Anthropic actually pauses a named training run citing RSP + RSI essay triggers.

Conf: H

11. P(RSP hard stop invoked by 2028) = 0.12

Claim: P(Anthropic publicly invokes RSP ASL hard stop with documented safeguard escalation by 2028) = 0.12.

Why: RSP v3.0 most detailed framework; v3.0 honestly splits commitments vs recommendations — weakens binding mass. Zero invocations 2024–26 despite Apollo scheming, Greenblatt misalignment reports. 12% = tail-gov adjacent, not modal.

Evidence:

  • (internal note) §23 Anthropic RSP v3.0
  • Node 4 §23 — P(RSP hard stop) = 0.12 modal
  • Anthropic RSP v3.0

Analogue: Fed “break glass” emergency lending — exists, rarely used pre-crisis.

Would update if: Misalignment Risk Report triggers ASL-4+ and public deployment delay >30d.

Conf: M

12. OpenAI Oct 2025 restructuring — paper governance ↑

Claim: OpenAI Foundation + OpenAI Group PBC recap (Oct 28, 2025) increases nominal safety governance (SSC, mission-only fiduciary on safety) but does not yet lower P(no halt) — track record discount applies.

Why: Structure after superalignment gutting and before proof of SSC veto use. CA AG MOU conditions non-objection to recap. Kolter SSC chair exclusive Foundation board — separation on paper.

Evidence:

Analogue: Post-2008 bank living wills — compliance document until tested.

Would update if: SSC blocks a named model release with public CA AG notice within 12 mo.

Conf: M

13. P(OpenAI SSC blocks frontier release by 2028) = 0.15

Claim: P(OpenAI Safety and Security Committee blocks or delays a frontier model release past internal schedule by 2028) = 0.15.

Why: CA MOU grants SSC role over release when risk thresholds exceeded; higher than P(full training halt) because deployment veto is narrower than compute stop. Still no observed block 2025–26 post-recap.

Evidence:

  • CA AG MOU § — SSC advance notice, material risk
  • OpenAI structure page — SSC “continue its current role” over safety practices
  • Node 11 lab tracker — OpenAI P(halt|crisis) = 0.08 vs P(SSC block) = 0.15

Analogue: FDA complete response letter — delays launch, not factory shutdown.

Would update if: GPT-5.x or successor launch publicly delayed citing SSC.

Conf: L–M

14. Anthropic LTBT board majority — Apr 2026

Claim: LTBT-appointed directors achieving board majority (Narasimhan, Apr 14, 2026) is real governance power on paper but P(invoked to delay release) ≈ 0.12 given Sharma exit + RSI continue-implied.

Why: Trust holds Class T shares; trustees no equity — mission alignment by design. Majority first time — but concurrent insider exodus and expect not commit pause rhetoric.

Evidence:

Analogue: Dual-class founder control — potential veto, observed product speed.

Would update if: LTBT board publicly blocks DoD contract or frontier deploy citing public benefit.

Conf: M

15. P(Anthropic conditional pause → actual multilab halt) = 0.08

Claim: P(Anthropic RSI-style conditional pause ** converts to** ≥2-lab verifiable halt by 2028) = 0.08.

Why: Requires (a) verification infrastructure built, (b) other frontier labs sync, (c) Anthropic acts not expects. Node 3 P(verified training-limit treaty) ≈ 0.04–0.06 — upper bound on multilateral halt.

Evidence:

  • Anthropic_When_AI_Builds_Itself_2026.md §1.1 — four conditions
  • Node 9 — P(durable multilateral pause) < 0.05
  • Node 4 §219 — “No revenue-peak voluntary halt precedent”

Analogue: Paris climate — conditional pledges without enforcement.

Would update if: Anthropic Institute publishes working verification protocol adopted by OpenAI + DeepMind.

Conf: M

16. P(whistleblower pipeline thins 2026–28) = 0.60

Claim: P(insider advocacy shifts from organize/leak to exit/ silence over 2026–28) = 0.60.

Why: Sharma pattern + Kokotajlo costly signal rare + Right to Warn transparency not pause + NDA enforcement. E4 managed disclosure substitutes for leak.

Evidence:

  • Node 4 §105 — “Safety employee pipeline structurally homeless post-Sharma”
  • Node 4 §162 — insider exhaustion ↑
  • (internal note) §6.4 — exit派 vs 中悲观派 table

Analogue: Post-Snowden tech — fewer new whistleblowers, more internal compliance theater.

Would update if: ≥3 named frontier researchers testify under oath with document dumps in 12 mo.

Conf: M

17. P(safety org reports to product/revenue, ≥2 labs by 2028) = 0.45

Claim: P(at least two frontier labs have safety/research integrity org formally reporting to CPO, product, or revenue lead by 2028) = 0.45.

Why: OpenAI post-superalignment absorption documented; Google Apr 2024 moved Responsible AI into DeepMind model-building (Building for our AI future). Industry trend = proximity to models framed as safety; independence ↓.

Evidence:

  • (internal note) §23 OpenAI — “absorbed into other divisions”
  • Google blog Apr 2024 — Responsible AI → DeepMind
  • Crosscut §2 would-update — “safety reorg reporting to product/revenue” ↑ hollowing

Analogue: Security embedded in DevOps (DevSecOps) — faster shipping, weaker veto.

Would update if: OpenAI or Anthropic publicly restores CEO-only reporting line for Alignment/Safeguards with board seat.

Conf: M

18. Google DeepMind — AI Safety org + FSF v3.1

Claim: DeepMind framework sophistication ↑ (FSF v3.1 Apr 2026, AI Safety & Alignment org Feb 2024) but P(FSF triggers deployment pause) ≈ 0.10 — self-enforced, Alphabet product pressure.

Why: FSF adds TCLs, SL2+, misalignment domain merged with ML R&D; governance section describes escalation to “appropriate corporate governance bodies” — no public halt precedent. Gemini launch cadence continues.

Evidence:

Analogue: OpenAI Preparedness — sophisticated paper, untested brake.

Would update if: Gemini frontier launch delayed citing FSF CCL threshold.

Conf: M

19. Meta / xAI — weakest governance; race pressure ↑

Claim: P(Meta or xAI voluntary halt | internal crisis) < 0.03; P(xAI mission drift under defense/commercial pressure by 2028) = 0.75.

Why: Meta — open-weight philosophy, no RSP-class hard stop. xAI — Feb 2026 Pentagon classified Grok deal “all lawful purposes” vs Anthropic safeguards refusal narrative; no formal alignment framework.

Evidence:

  • (internal note) — xAI Pentagon Feb 2026
  • (internal note) §23 Meta — weakest formal framework 4/10
  • Crosscut §2 — “xAI all lawful purposes Pentagon deal vs Anthropic safeguards”

Analogue: Open-source vs safety — recall impossible once weights out.

Would update if: xAI publishes RSP-equivalent with external audit.

Conf: M

20. P(E4 managed disclosure preempts halt pressure) = 0.40

Claim: P(when internal alignment concern rises, lab publishes managed longform (RSI essay class) instead of leak + continues training) = 0.40.

Why: Anthropic Institute model — own narrative before whistleblower. Reduces costly leak incentive; institutionalizes transparency-without-brakes (Cluster D).

Evidence:

  • Node 4 §32, P(E4)=0.37
  • Anthropic_When_AI_Builds_Itself_2026.md §5.2 — company framing + conditional pause
  • Node 4 §417 — “We take this seriously + continue”

Analogue: Corporate pre-bunking in crisis PR.

Would update if: Managed disclosure immediately followed by halt (never observed).

Conf: M

21. Revenue / capex peak prevents voluntary halt P = 0.70

Claim: P(frontier labs do not voluntarily halt during Stargate-class capex / revenue peak, conditional on governance crisis) = 0.70.

Why: Node 1 labor/capex modal + Node 4 “no revenue-peak voluntary halt precedent.” OpenAI/Microsoft/Google $100B+ AI infra commitments 2025–26 — opportunity cost of halt ↑.

Evidence:

  • Node 4 §161 — “Revenue peak + capex cycle — no voluntary halt precedent”
  • Node 1 — agent labor shock modal reinforces commercial default
  • OpenAI recap Oct 2025 — investor returns + IPO path (NBC)

Analogue: Oil majors during price spike — delay climate capex, not stop pumping.

Would update if: Major lab publicly cuts training budget >20% citing safety (not market).

Conf: M

22. PBC mission clause invoked to delay release P = 0.10

Claim: P(any frontier PBC publicly cites mission/benefit clause to delay a frontier release 2026–28) = 0.10.

Why: Legal latitude exists (OpenAI PBC articles, Anthropic charter); observed invocations zero. Clauses function as litigation shield and marketing, not operational brake.

Evidence:

  • OpenAI PBC incorporation — safety-only fiduciary language
  • Anthropic PBC public benefit purpose (research_ai_governance_landscape.md)
  • Crosscut §2 — “PBC erosion under revenue/capex pressure”

Analogue: B-Corp certification — rarely blocks profitable product.

Would update if: SEC or CA filing cites PBC duty as reason for delayed launch.

Conf: L–M

23. P(second OpenAI governance crisis by 2028) = 0.18

Claim: P(OpenAI experiences second C-suite/board governance crisis with global media coverage by 2028) = 0.18.

Why: First crisis 2023; recap concentrates power in Foundation but same CEO; investor/IPO pressure + safety tensions recur. Not identical to P(halt) — crisis can fire without halt (T1 branch).

Evidence:

  • 2023 crisis base rate — 1 event in ~4 yr company history at frontier scale
  • Oct 2025 structure — Altman remains CEO on both boards
  • AI_safety_大事记 — “安全 vs 商业” recurring narrative

Analogue: Uber 2017 — repeated governance scandals, no business model halt.

Would update if: 24 mo without any OpenAI governance headline (downward).

Conf: L

24. Superalignment arc repeats at another lab P = 0.35

Claim: P(another frontier lab ** announces** dedicated superintelligence/safety team with compute pledge, then dissolves or hollows it within 24 mo, 2026–28) = 0.35.

Why: DeepMind AI Safety & Alignment (2024) parallels superalignment narrative; industry pattern = announce → absorb. OpenAI template copied then abandoned.

Evidence:

  • TechCrunch 2024 — DeepMind org “Similar in mission to Superalignment”
  • OpenAI superalignment arc (AI_safety_大事记 §3.2)
  • research_ai_governance_landscape.md — pattern of contradiction

Analogue: Corporate innovation lab lifecycle — launch, hype, merge into core product.

Would update if: DeepMind maintains separate AGI safety team with published budget through 2028.

Conf: L–M

25. Node 11 → P(no effective pause) +3–5pp

Claim: Integrating Node 11 (sim) raises P(no federal/industry training pause through 2028) from 0.82–0.84 to 0.85–0.89 (+3–5pp).

Why: 65% modal hollowing + 25% crisis-without-halt both imply continue; only 10% re-empower tail reduces. Confirms Node 4 modal whistleblower branch structurally underpinned.

Evidence:

  • (internal note) crux #5 — 0.82 (0.72–0.92)
  • Node 9 net on P(no pause) +0.02–0.03 — Node 11 stronger channel (US labs)
  • Crosscut §2 p(doom) — “+3–5pp vs −2–3pp” on extinction from hollowing vs re-empower

Analogue: Bayesian — same evidence, explicit institutional prior.

Would update if: Re-stitch with invoked RSP hard stop observed.

Conf: M

26. Node 11 → P(no shutdown at ASI) +3–5pp

Claim: Node 11 raises P(no effective shutdown when ASI threshold reached) from ~0.60 to ~0.63–0.65 (+3–5pp).

Why: Labs that cannot halt at C7–C9 (governance evidence) unlikely to execute shutdown at C10 without external coercion (DPA 0.03, EU deployment 0.14 — insufficient for training stop). Node 4 multiplier 3 chain.

Evidence:

  • Node 4 §672 — RSP hard stops 0.12; if not credible, raises P(no shutdown)
  • (internal note) — failure mode bang vs whimper; modal = oversight not halt
  • Node 4 §585 — P(no effective shutdown at ASI) × 0.50–0.70 in extinction chain

Analogue: Fukushima — institutions fail to SCRAM when design assumes they will.

Would update if: Live-fire shutdown drill at ≥1 frontier lab with public report.

Conf: M

27. Cluster A correlation — hollowing ↔ race ↔ no pause

Claim: P(modal hollowing) correlates ↑ with P(no pause), P(race post-theft N3), P(GAAIA preemption N6) — joint tail ~3–5× product of marginals, not independent.

Why: Same capital/capex/natsec coalition blocks halt. Node 11 makes explicit what correlation matrix listed for Cluster A.

Evidence:

  • (internal note) Cluster A — acceleration block
  • Crosscut §2 links table — N1, N3, N4, N6
  • Node 3 — post-theft race correlates with hollowing

Analogue: Correlated defaults in CDOs — joint tail risk.

Would update if: EU-only brake decouples US lab pace from governance (Node 9 T1 tail).

Conf: M

28. p(doom) net — extinction bucket +1.5–3pp

Claim: Node 11 contributed ~+1pp to Phase 2b re-stitch (prior pass → ~17% doom region), primarily via misalignment tail branch (+0.5–1pp) and modal race speed (+0.5pp).

Why: Does not add independent bio risk. Misalignment branch weight 25% × higher P(no shutdown) × continued training → compound not additive — use mixture not product of 28 marginals.

Evidence:

  • (internal note) — branch model 58% modal / 25% misalign / 17% bio
  • Crosscut §2 — “+3–5pp on P #1 and P #6”
  • Node 11 parent doc — p(doom) table

Analogue: Phase 4a stitch methodology — node-by-node delta.

Would update if: Full re-stitch with correlation_matrix Cluster A discount.

Conf: L — pending Phase 4b calibration

Update log

DateChange
2026-07-04Initial — 28 evidenced P sections