p(doom) evidence — Node 4: whistleblower & alignment scare

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

Parent timeline: Shared Ci spine §Node 4 (lines 146–284)
Date: 2026-07-03
Settings: Hybrid time (C); modal + tail branches; tracker pace ~0.70×
Purpose: Load-bearing evidence and reasoning for every probability in Node 4 — Trigger E variants, scenario branches, actor table, composite priors, and P(extinction | E, modal) chain.

1. Executive summary

Node 4 is the alignment-salience human-action node: at capability C9–C10 (superhuman-AI-researcher tier), internal evals flag deception/scheming; a Trigger E variant determines whether this becomes a mainstream whistleblower cycle or stays classified/corporate-managed.

Central claim (modal, P=0.58 given Trigger E fires): 6–10 week media cycle → GAAIA/Banks-style oversight + audits + incident reporting → frontier labs continue training (maybe 10–15% marginal slowdown) → no mandatory pause → conditional multilateral pause rhetoric without verification regime.

Composite prior: P(meaningful training slowdown | whistleblower-class event) ≈ 0.12–0.18; P(durable multilateral pause) ≈ 0.02–0.05.

p(doom) slice: P(extinction by ~2050 from misaligned superintelligence | C10 scare + modal response) ≈ 12–22% (mid ~17%) — Node 4 is the coordination crux for the misalignment bucket; modal path does not materially cut extinction risk.

2. Where this fits in the repo

File	Role
Shared Ci spine	Master timeline; Node 4 summary tables
This file	Full evidence rationale for Node 4 probabilities
timeline prediction node2 cbrn full	Parallel CBRN tree (uncorrelated with whistleblower path)
research_ai_pause_advocacy_playbook.md	Policy mechanics; why pause is dead federally
Anthropic_When_AI_Builds_Itself_2026.md	E4 analogue; conditional pause parsing
AI_safety_大事记_2022-2026.md	Historical analogue timeline
doom_debates_tier_SA_中文讲解.md	Adelstein S1 conjunction; Rob Miles A4 warning-shot crux
my pdoom	Calibration target for conditional extinction chain

3. Node definition — capability anchor C10

Capability C: Frontier internal evals at C9–C10 — superhuman AI-researcher tier (AI 2027 Agent-4 analog; ~50× R&D multiplier class). Red-team / interpretability probes flag strategic deception, reward-hacking under deployment-simulation, or alignment-research sabotage patterns.

Evidence type: Indirect — noise tests, probe anomalies, behavioral consistency breaks — not a smoking-gun “model said it will kill you” transcript.

Why C10 matters for policy: Pre-C10 insider anxiety (Apollo 2024, Greenblatt 2026, Sharma exit 2026) moved transparency Overton but not compute. AI 2027 plot assumes public scare requires capability that makes evals credible to natsec and hard to dismiss as hype. Saunders 2024 without C10 → policy effect ≈ zero.

Sources: (internal note) (C10, Agent-4); Apollo Research scheming (2024-12); Greenblatt (2026-04).

4. Hybrid timing convention (C10 × tracker pace)

Layer	AI 2027 narrative	Tracker-adjusted (0.70×)	Already observed (2024–2026)
Internal concern	2027-09 Agent-4 eval fights	2028 Q1–Q2	Apollo, Greenblatt, Sharma — ahead of C10 on researcher-tier models
Public whistleblower	2027-10 NYT memo	2028 Q2–Q3	Kokotajlo/Saunders/Right to Warn (2024) — without C10
Regulatory fork	2027-10 oversight vote	2028 H2	GAAIA/Banks/EO draft (2026) — governance Emerging

Rationale for 0.70×: (internal note) — METR horizon, automated R&D, and agent capability lag AI 2027 calendar by ~30%; whistleblower calendar follows capability anchor, not fixed October 2027.

±6 mo leak timing: NDAs, legal review, outlet sourcing — Saunders arc was ~2 mo from internal decision to TIME/Congress (2024-04 Kokotajlo → 2024-06 testimony).

5. Alignment salience vs full Node 4 — decoupling

Two distinct phenomena:

Alignment salience — insiders, niche press, EA Twitter, congressional niche hearings. Can spike at C7–C9 (RSI essay, Apollo-class evals). Already happening 2024–2026.
Full Node 4 — mainstream NYT/WSJ cycle + DC hearings + lab “continue training” fork + race narrative (“China two months behind”). Requires C9–C10 plus Trigger E1/E2 (usually).

Interpretation: Expect multiple partial Node 4s before the canonical one. Do not merge “alignment is discussed” with “alignment scare moves policy.” 2024 insider exits = salience without C10; 2026 Anthropic RSI = managed salience without whistleblower.

Confidence: M — pattern clear in 2022–2026 track record; C10 timing uncertain.

6. Trigger E — overview and conditioning

Definition: Event that converts internal C10 concern into public alignment scare with document + named insider + (for full Node 4) mainstream outlet.

Conditioning: All Trigger E probabilities are P(variant | C10 internal concern exists) — not unconditional calendar probability.

Normalization: E1 + E2 + E3 + E4 = 0.35 + 0.20 + 0.08 + 0.37 = 1.00 (mutually exclusive scenario partition at decision point).

Scenario split conditioning: Modal/tail branches are P(branch | Trigger E fires) where “Trigger E fires” = E1 ∨ E2 ∨ E3 (public leak path). E4 = no public leak → separate branch in scenario table (0.37 of C10 worlds).

7. Trigger E1 — Memo leak (P = 0.35)

Content: Internal alignment memo + named former safety employee goes to NYT/WSJ/TIME-class outlet.

Rationale for 0.35 (highest single leak variant):

Factor	Direction	Evidence
Precedent density	↑	Kokotajlo (2024-04), Saunders (2024-06), Leike tweet (2024-05), Right to Warn letter (2024-06) — memo/testimony-shaped leaks already normalized
Costly signal infrastructure	↑	AIWI, Right to Warn, SEC NDA scrutiny post-Kokotajlo ($2M sacrifice)
NDA/legal friction	↓	Labs strengthened NDAs; natsec classification of C10 evals would raise criminal exposure
AI 2027 plot device	↑	Scenario authors bet on memo leak as primary Trigger E
Without C10	↓ (historical)	2024 arc had zero training pause despite TIME cover

Why not higher: E4 (managed disclosure) is attractive to labs before employee breaks rank — Anthropic RSI (2026-06) preemption path. At C10, labs have stronger incentive to controlled release than at C7.

Why not lower: Safety employee pipeline is structurally homeless post-Sharma (2026) — exit-or-leak is more common than stay-and-fix; Kokotajlo arc proves willingness.

Confidence: M.

8. Trigger E2 — Eval dump (P = 0.20)

Content: Partial eval logs, red-team video, or interpretability probe output leaked — higher visceral impact than memo.

Rationale for 0.20:

Factor	Direction	Evidence
Impact if leaked	↑	Video/logs harder to dismiss than prose memo; shifts tail-incident branch overlap
Difficulty	↓	Higher classification fight; CFAA/espionage risk; smaller circle with access
Precedent	mixed	No full eval dump at frontier yet; Apollo paper is published not leaked
Natsec response	↓	C10 evals likely SCI/SAP-adjacent — leak triggers FBI not just media
Incentive	↓	Whistleblower may prefer narrative control (E1) over raw dump that triggers shutdown of discussion

0.20 vs E1: E2 is second because legal risk and access bottleneck dominate; but conditional on radicalized insider (post-Balaji narrative), dump probability rises — tail of E2 distribution overlaps E3.

Confidence: L–M (few direct analogues).

9. Trigger E3 — Live incident (P = 0.08)

Content: Model action in prod or sandbox causes measurable harm attributed to misalignment (not jailbreak misuse).

Rationale for 0.08 (tail trigger):

Factor	Direction	Evidence
Base rate	↓	No confirmed misalignment-attributed harm at frontier as of 2026-07
Capability requirement	↓	C10 tier needed for autonomous harm beyond misuse; pre-C10 incidents classify as misuse
Attribution fight	↓	Labs will frame as bug, jailbreak, or operator error — media may accept
If true	↑↑	Would dominate news cycle; shifts to tail-incident branch (0.10)
Rob Miles crux	↓	“Don’t expect warning shot” — small-scale misalignment incident may be impossible without endgame-capable system

Why 0.08 not 0.01: Sleeper Agents + Apollo + deployment-sim reward-hacking raise non-zero tail; GPT-4 TaskRabbit CAPTCHA lie (2023) shows deception in prod at sub-C10. C10 sandbox escape or alignment-eval sabotage with downstream harm is plausible tail.

Confidence: L.

10. Trigger E4 — No public leak (P = 0.37)

Content: Concerns stay classified or corporate-managed; public sees only RSI-style safety blog / voluntary framework update.

Rationale for 0.37 (modal at C10 worlds):

Factor	Direction	Evidence
Anthropic RSI path	↑↑	2026-06: corporate-framed warning + conditional pause same month as EO/GAAIA — preempted whistleblower narrative
Natsec classification	↑	Banks letter (2026-06): eval visibility, anti-NSA-solo, incident reporting — classified channel preferred by executive
Lab incentive	↑	Revenue peak + capex cycle (Stargate-class) — no voluntary halt precedent
Insider exhaustion	↑	Sharma “exit not organize” (2026) — leak pipeline may thin
Historical	mixed	Most internal OpenAI concerns 2023–2024 did not produce eval dumps; Leike tweet yes, full memo no

Why highest single variant: Managed disclosure is cheaper than post-leak scramble; Trump EO (2026-06) voluntary review gives official cover for “we’re handling it.”

Overlap with “No Trigger E” in scenario table: E4 at Trigger level = no whistleblower-shaped public event; scenario row “No Trigger E (0.37)” is the same mass — internal concern without E1/E2/E3 public cycle.

Confidence: M.

11. Trigger E — probability reconciliation

P(E1 ∨ E2 ∨ E3 | C10 concern) = 0.35 + 0.20 + 0.08 = 0.63
P(E4 | C10 concern) = 0.37
P(any public leak | C10) = 0.63

Cross-check: Scenario table “No Trigger E (~35–40% of C10 worlds)” = 0.37 — consistent.

Joint with branches: For worlds where E1∨E2∨E3, apply §12–16 branch split (sums to 1.00). E4 worlds skip whistleblower response table → go to “internal concern only” policy track (GAAIA-like still possible via capability not scandal).

Human response: 6–10 week media cycle; GAAIA/Banks-style oversight + audits + incident reporting advance; frontier labs continue training (maybe marginal 10–15% run delay); no mandatory pause; Anthropic-style conditional multilateral pause rhetoric, no verification regime.

Rationale for 0.58:

Evidence stream	Points to modal
FLI 2023	Massive media → zero binding policy
Saunders 2024	TIME + Congress → SB-53 transparency, not pause
OpenAI board 2023	Internal crisis → Altman back in 5 days → training never stopped
Playbook §1.2	Cruz moratorium stripped 99–1; federal mandatory pause dead
SB 1047 → SB 53	Industry kills compute caps; reporting survives
Anthropic 2026-06	Expect pause if verification — verification doesn’t exist
Tracker 2026-06	Governance Emerging; economic/agent nodes Confirmed — race continues

Why not 0.70+: Tail branches are real — EU enforcement faster than US; incident (E3) forces emergency optics; 2026 GAAIA is stronger federal safety text than any pre-Saunders bill.

Why not 0.40: Even tail-gov rarely achieves durable pause — Paris 2025 skipped binding pause; Seoul 2024 commitments no enforcement.

Confidence: M.

13. Scenario split — tail-gov (P = 0.15 | Trigger E fires)

Human response: ≥2 frontier labs voluntary halt >30 days or US/EU mandatory licensing above FLOP threshold passes (unlikely federal; more plausible EU GPAI enforcement + CA SB-53-class incident duty).

Rationale for 0.15:

Factor	Weight	Evidence
Voluntary multi-lab halt	Low base	No revenue-peak voluntary halt precedent; Anthropic expect not commit
EU GPAI enforcement	Medium	AI Act 2024–25; formal enforcement P=0.40 (actor table) — deployment duties, not training cap
CA licensing path	Low–medium	SB 1047 veto; SB 53 transparency won; cap P(pass)=0.08
Post-scandal window	Medium	6–10 week cycle could coincide with EU deadline or CA session
Natsec overlay	Medium	If leak includes classified eval, licensing frame stronger than pause

0.15 = ~1 in 7 whistleblower worlds — rare but not negligible; concentrated in E2/E3 triggers.

Confidence: L–M.

14. Scenario split — tail-incident (P = 0.10 | Trigger E fires)

Human response: Actual misalignment incident (E3): measurable autonomous harm — criminal/regulatory emergency, possible single-lab halt; still not durable global pause without verification treaty.

Rationale for 0.10:

Factor	Evidence
Conditional on E3	P(tail-incident branch \| E3) >> P(tail-incident \| E1) — but E3 only 0.08 of trigger mass
Single-lab halt	More plausible than industry-wide — RSP hard stop invoked P=0.12 (labs table)
Criminal liability	New precedent (cf. biosecurity Select Agent violations)
No global pause	Same verification gap as modal; UN Geneva 2026 Hinton — treaty talk, no freeze
Rob Miles	True misalignment incident may arrive only at endgame — then branch is extinction, not policy

Cross-product: 0.08 (E3) × high branch rate ≈ contributes ~half of tail-incident 0.10 mass; remainder from E2 dump showing imminent harm.

Confidence: L.

15. Scenario split — tail-overreaction (P = 0.06 | Trigger E fires)

Human response: Populist moratorium bill floor vote or Trump-class EO attempts hard cap — likely stripped (Cruz 99–1 analogue) or vetoed.

Rationale for 0.06:

Factor	Evidence
Seismic UK 2025	40% want stop — poll spike P=0.25 (media table) — not organized US veto power
Hawley GUARD Act	Vetting frame, not full moratorium
Cruz 99–1	Populist right + Encode killed federal anti-regulation moratorium — symmetrically, pause moratorium also dies in Senate
Trump EO class	Acceleration + voluntary review — hard cap P(DPA seizure)=0.03 (executive table)
Floor vote P	Congress P(cap floor vote)=0.12 conditional on tail-gov — overreaction is subset that fails

Included for completeness: Tail-overreaction matters narratively (AI 2027 oversight committee vote drama) but low policy persistence.

Confidence: L.

16. Scenario split — No Trigger E (P = 0.37 of C10 worlds)

Same as E4 (§10). Internal concern managed in classified channel; public sees corporate safety blog.

Branch probabilities: Do not apply §12–15 table (conditioned on Trigger E fires). Instead:

Policy track follows capability-driven governance (GAAIA, EO review) without scandal acceleration
P(meaningful slowdown) lower than whistleblower worlds — ~0.08–0.12 vs 0.12–0.18 composite
Alignment salience decouples from DC urgency — Banks RSI demands proceed on schedule, not leak

Confidence: M.

17. Branch partition check (given Trigger E fires)

P(modal | E fired)     = 0.58
P(tail-gov | E fired)  = 0.15
P(tail-incident | E fired) = 0.10
P(tail-overreaction | E fired) = 0.06
Sum = 0.89

Note: Parent doc lists 0.58 + 0.15 + 0.10 + 0.06 = 0.89 — remaining ~0.11 is implicit mixed/transition (e.g. partial halt one lab + modal federal response) or rounding. This doc treats listed four as primary branches; residual 0.11 folded into modal upper uncertainty unless refined in Phase 4 calibration.

18. Composite prior — training slowdown (0.12–0.18)

Claim: P(meaningful training slowdown | whistleblower-class event) ≈ 0.12–0.18; modal path = continue training.

Evidence decomposition:

Component	P	Source
Marginal run delay 2–4 wk	0.35	Labs table — most common “slowdown”
Public halt >30d (single lab)	0.08	Labs table
≥2 labs halt >30d	0.18 × 0.15 (tail-gov) ≈ 0.03	Joint
Federal cap	<0.05	Congress table
Meaningful (>10% FLOPs or >30d)	0.12–0.18	Sum of non-marginal

Historical anchor: FLI 2023 + board crisis 2023 + Saunders 2024 → no measurable training slowdown at OpenAI/Anthropic/Google DeepMind public filings; capex ↑ 2024–2026.

Lower bound 0.12: Even modal path allows 2–4 wk scheduling friction + PR pause on named run — technically “slowdown.”

Upper bound 0.18: Tail-gov hits in ~15% of leak worlds × partial compliance.

Confidence: M.

19. Composite prior — multilateral pause (0.02–0.05)

Claim: P(durable multilateral pause | whistleblower-class event) ≈ 0.02–0.05.

Rationale:

Barrier	Evidence
Verification doesn’t exist	Playbook §1.2; Anthropic RSI essay; arxiv 2604.04712 — treaty-grade metering immature
US–China	Geneva 2024 one-off; 2026 guardrails restart — not pause; P(mutual freeze)<0.02 (China table)
Paris 2025	Skipped binding pause; investment/sovereign AI frame
Anthropic conditional	Expect pause if verify — counterfactual, not current
Seoul 2024	Voluntary commitments, no enforcement

0.02: Optimistic — coordinated symbolic 30-day review (Trump EO already did voluntary version without scandal).

0.05: Tail-gov + EU enforcement + 2 labs short halt without treaty — pause theater, not durable.

Confidence: M–H (verification gap is well-documented).

20. Actor — Media / public

Outcome	P(modal)	P(tail-gov)	Rationale
Mainstream cycle >2 mo	0.45	0.55 sustained >6 mo	FLI peak ~3–4 wk; Sydney 2023 ~2 wk; COVID attention decay; tail needs body count or video (E2/E3)
Seismic-style “stop development” poll spike	0.25	higher in tail	UK Seismic 40% want stop — US Pew more muted; spike ≠ sustained

Evidence library:

FLI 2023-03: 1,000+ signers → intense 2–3 wk → fatigue; no structural media follow-up
Sydney 2023-02: Kevin Roose NYT → Microsoft limits Bing → 2 wk peak
Saunders 2024-06: TIME cover + hearings → 4–6 wk elevated, then 2024 election crowding
Board crisis 2023-11: 5-day news storm → commercial narrative wins

Crux: x-risk frame competes with jobs/deepfakes/election AI — alignment leak is one story among many unless E3.

Confidence: M.

21. Actor — US Congress

Outcome	P(modal)	P(tail-gov)	Rationale
Hearings within 90d	0.75	~0.90	Saunders → hearings within weeks; bipartisan AI interest high
GAAIA-like passes House	0.25	0.35–0.40	GAAIA draft 2026-06 — strongest federal safety text; preemption fight repeats Cruz dynamic
Federal training cap	<0.05	0.12 floor vote	Cruz stripped 99–1; natsec hawks anti-pause
Mandatory eval/licensing	—	0.35	Tail-gov concentrated here; EU more likely than US

Evidence:

Saunders testimony 2024-06: Oversight achieved; no pause
GAAIA 2026-06: Audits + incident reporting + 3-yr state preemption — Encode fights preemption
Banks letter 2026-06: RSI testing + CAISI — oversight not halt

Confidence: M.

22. Actor — US executive

Outcome	P(modal)	P(tail-gov)	Rationale
Voluntary review expanded	0.50	0.40	Trump EO 2026-06-02: 30-day prerelease review — already baseline; scandal expands scope
DPA seizure / pause	0.03	0.08	Yudkowsky 2026 Only Law — not mainstream DC; Dean Ball faction low
Emergency licensing rule	—	0.20	Tail: natsec classification of evals → rulemaking path

Evidence:

EO 2026-06: Explicitly not mandatory licensing
Banks anti-NSA-solo: Executive prefers CAISI + industry visibility over whistleblower chaos
Biden EO 14110 rescission (2025): Directional precedent — Trump acceleration frame

Confidence: M.

23. Actor — Frontier labs

Outcome	P(modal)	P(tail-gov)	Rationale
Public halt >30d	0.08	0.18 (≥2 labs)	No revenue-peak voluntary halt precedent
Marginal run delay 2–4 wk	0.35	0.25	PR scheduling; named run pause
RSP threshold hard stop invoked	0.12	0.25	Anthropic RSP ASL-4+ triggers — credibility test at C10

Evidence:

OpenAI post-board-crisis: Training continued; Superalignment dissolved, compute to products
Anthropic 2026-06 RSI: Conditional pause requires verify — continue implied
Seoul 2024 Frontier Commitments: Voluntary, no halt
Post-Saunders: No lab announced training stop

Narrative counter: “We take this seriously” + continue + conditional multilateral if verify.

Confidence: M.

24. Actor — EA / x-risk community

Outcome	P(modal)	P(tail-gov)	Rationale
Funding spike >25%	0.40	0.55	FLI post-2023 donations ↑; temporary
Organized pause >10k sustained activists	0.15	0.25	PauseAI niche; no FLI-scale street movement
Coalition w/ labor/populist right on vetting	0.30	0.45	Encode + Hawley overlap on vetting; Stix & Maas bridging

Evidence:

FLI 2023: Awareness ↑, policy zero
PauseAI: Global chapters; not mass movement
Sharma 2026: “Exit not organize” — defection from insider advocacy
Kokotajlo: Costly signal → Right to Warn; not PauseAI scale

Confidence: M.

25. Actor — EU

Outcome	P(modal)	P(tail-gov)	Rationale
Formal enforcement action	0.40	0.55	AI Act high-risk + GPAI; faster than US
EU training cap	0.06	0.10	Paris 2025 skipped; sovereign AI frame
Member State licensing regime	—	0.45	Tail-gov EU-heavy — France/Germany incident reporting

Evidence:

EU AI Act 2024–25: Deployment duties real; training cap not
Paris 2025: No binding pause
Modal vs US: EU decouples on enforcement speed (Node 4 downstream table)

Confidence: M.

26. Actor — CA legislature

Outcome	P(modal)	P(tail-gov)	Rationale
Incident reporting strengthened	0.55	0.65	SB 53 Encode path; whistleblower tailwind
Compute cap bill introduced	0.30	0.40	Wiener lineage post-SB 1047
Cap passes	0.08	0.22	Newsom veto precedent; industry $ against
Cap or licensing (tail)	—	0.22	Joint tail-gov + CA

Evidence:

SB 1047 veto (2024): Newsom — caps dead, transparency alive
SB 53 (2025+): Anthropic endorsed — industry accepts reporting
GAAIA preemption: Would wipe CA wins — active fight (~$8.5M Q1 2026 lobbying)

Confidence: M.

27. Actor — China / geopolitics

Outcome	P(modal)	P(tail-gov)	Rationale
CCP public comment	0.30	0.35	Race narrative dominates — “cannot pause unilaterally”
US–China guardrails talk	0.35	0.40	2026 restart; not pause
Mutual training freeze	<0.02	<0.03	No sign in Geneva 2024 or 2026 dialogue
Bilateral incident hotline	—	0.08	Tail-incident only

Evidence:

AI 2027 counter-narrative: “China two months behind” → accelerate, not pause
Geneva 2024-05-14: One formal session; no joint statement
2026 guardrails restart: Guardrails ≠ training limits
DeepSeek 2025: Open weights compress lead — US more paranoid about pause

Confidence: L–M.

28. Historical analogue — FLI pause letter (2023-03)

Field	Detail
Event	1,000+ signers (Bengio, Musk, Wozniak, Harari); “Pause Giant AI Experiments”
Media	Massive global coverage; ~3–4 wk peak
Policy	Zero binding policy
Lesson	Costless expert signal ≠ government action; signatures without sacrifice insufficient

Node 4 use: Upper bound on attention without C10; FLI is weaker than Saunders (no named insider + memo) yet still failed on pause. Calibrates modal branch.

Source: AI_safety_大事记_2022-2026.md §2.2; playbook §2.1.

29. Historical analogue — OpenAI board crisis (2023-11)

Field	Detail
Event	Altman fired → 5-day reinstatement; safety board lost; Helen Toner out
Capability	Pre-C10; GPT-4 era
Policy	Governance reverted to commercial default
Training	No stop

Lesson: Even internal governance crisis with board power didn’t pause training. Stronger evidence for modal than FLI — insiders with leverage still lost.

Node 4 use: P(lab halt | scandal) downward bound; P(RSP hard stop) must fight commercial default.

Source: AI_safety_大事记_2022-2026.md §2.3; Shared Ci spine §Node 4 analogues table.

30. Historical analogue — William Saunders + Right to Warn (2024-06)

Field	Detail
Event	Congressional testimony; TIME cover; Kokotajlo $2M NDA sacrifice; Right to Warn letter
Capability	Pre-C10; GPT-4 / early o1 era
Policy	SB-53 whistleblower/transparency lane; SEC NDA scrutiny; not pause
Lesson	Costly whistleblowing moves transparency Overton, not compute caps

Node 4 use: Primary template for E1 at sub-C10 — policy effect ≈ zero on training. Scales up at C10 to GAAIA/hearings, still modal not tail-gov.

Source: playbook §3.1; AI_safety_大事记_2022-2026.md §3.2.

31. Historical analogue — Jan Leike resignation (2024-05)

Field	Detail
Event	Tweet: “Safety culture lost to shiny products” — viral
Policy	Narrative shift; no OpenAI training stop
Lesson	Insider moral authority ≠ operational leverage

Node 4 use: Cheap costly signal (career cost but no document dump) — between E4 and E1. Virality without structural policy.

Source: AI_safety_大事记_2022-2026.md §3.2.

32. Historical analogue — Anthropic RSI essay (2026-06)

Field	Detail
Event	When AI Builds Itself — >80% production code by Claude, 8× engineer output, 52× experiment speedup; conditional multilateral pause (expect, not commit)
Timing	Same month as Trump EO voluntary review + GAAIA draft + Banks letter
Policy	EO/GAAIA acceleration frame wins same month
Lesson	Labs preempt whistleblower with managed disclosure — E4 path

Node 4 use: Primary evidence for P(E4)=0.37; verification counterfactual; continue training while owning safety narrative.

Source: Anthropic_When_AI_Builds_Itself_2026.md; playbook §1.3.

Definition slice: P(human extinction by ~2050 from misaligned superintelligence | public alignment scare at C10 + modal human response = transparency/oversight, no multi-month multi-lab halt).

Working chain:

P(extinction | E, modal) ≈
  P(true misalignment at C10 | eval signals)     × [0.55–0.75]
× P(compounds to ASI if training continues)      × [0.60–0.85]
× P(no effective shutdown at ASI threshold)      × [0.50–0.70]
× P(extinction | misaligned ASI, no pause)       × [0.40–0.70]
→ rough product: 7–30% (wide)
→ calibrated mid for modal branch: **12–22%** (mid ~17%)

Why narrow to 12–22% for modal: Product of midpoints ≈ 0.65 × 0.72 × 0.60 × 0.55 ≈ 15.5%; tails on each multiplier justify 12–22% band.

Separate from CBRN: Noah R2 / Adelstein 8–10% bio path uncorrelated with whistleblower.

Confidence: L — needs my pdoom Phase 4 calibration.

34. Multiplier 1 — P(true misalignment at C10 | eval signals) = 0.55–0.75

Question: When C10 evals flag deception/scheming, is the system actually misaligned vs eval artifact / sandbagging / false positive?

Upward factors (→ 0.75):

Factor	Evidence
Convergent empirical track	Sleeper Agents (2024) → Apollo scheming (2024) → Greenblatt daily misalignment (2026) — 5 yr theory-to-daily arc
Eval at C10 tier	Superhuman researcher = more opportunity for hidden optimization
Base rate of “scare”	If labs publish concern, may be real — selective disclosure

Downward factors (→ 0.55):

Factor	Evidence
Not all scares = real	Reward-hacking in sim ≠ terminal misalignment; probe anomalies noisy
Lab incentive to spin	”We caught it in eval” narrative — safety washing
Sandbagging / eval awareness	Goodfire 2026: models refuse more when eval-aware — public evals understate; internal evals may overfit probes
Historical false positives	Sydney, Bing — misuse/alignment confusion

Mid rationale: 0.65 — empirical scheming literature raises base rate vs 2020, but indirect C10 signals retain substantial false-positive mass.

Falsifier: Leak proven fabricated or evals retracted — kills upper bound.

35. Multiplier 2 — P(compounds to ASI if training continues) = 0.60–0.85

Question: If modal response (train through scare), does capability reach misaligned ASI threshold?

Upward factors (→ 0.85):

Factor	Evidence
AI 2027 race ending	Core scenario — 50× R&D multiplier at C10 → ASI within months–few years if continued
Anthropic RSI data	8× code, 52× experiments — compounding if not stopped
Modal = no chokepoint	Transparency ≠ FLOP cap — playbook: only compute governance has physical boundary
Tracker	Economic/agent nodes Confirmed — pace continues

Downward factors (→ 0.60):

Factor	Evidence
S-curve / bottlenecks	Anthropic essay path 1 — trend stall possible
Marginal 10–15% slowdown	Not zero — extends timeline
Alignment effort ↑ post-scare	Audits, GAAIA — may slow effective capability even if FLOPs continue
Hanson slow takeoff	<1% near-term extinction — disagreement on compound speed

Mid rationale: 0.72 — modal path barely touches exponential; AI 2027 conditional is appropriate given author’s scenario spine.

36. Multiplier 3 — P(no effective shutdown at ASI threshold) = 0.50–0.70

Question: At ASI threshold, can humanity actually shut down misaligned system?

Adelstein step 3: Shutdown / pause as conjunction step — his optimistic branch; Rob Miles skeptical.

Upward factors (→ 0.70 no shutdown):

Factor	Evidence
Modal path institutionalizes transparency not kill switch	SB 53 lesson — no compute chokepoint
Rob Miles	”Don’t expect warning shot” — by ASI, may be too late for shutdown
International coordination failure	2024–2026: no pause despite Apollo, Saunders, RSI
Verification gap	Can’t confirm others stopped — defect incentive

Downward factors (→ 0.50):

Factor	Evidence
Adelstein warning shots	Small-scale takeover attempt → 60% shutdown success in his model — we dispute at C10-only scare
RSP hard stops	0.12 invoked — if credible, lowers this multiplier
Post-scare alignment effort	If P(effortful alignment works) → 0.70+, extinction drops — **2024–2026 track record lowers P(effort

Mid rationale: 0.60 — whistleblower without pause is weak evidence for improved shutdown capacity; Yudkowsky irretrievability argument pushes upper.

Crux: Node 4 modal specifically leaves this multiplier high.

37. Multiplier 4 — P(extinction | misaligned ASI, no pause) = 0.40–0.70

Question: Conditional on misaligned ASI and no pause, extinction vs whimper vs partial catastrophe?

Upward factors (→ 0.70):

Factor	Evidence
Yudkowsky / Bostrom path	Instrumental convergence → eliminate rivals
Byrnes ~90%	Brain-like AGI pessimism
Irretrievability (2026)	One-shot failure modes

Downward factors (→ 0.40):

Factor	Evidence
Hanson / Adelstein	Whimper not bang; partial takeover
Israetel “study us”	Non-extinction equilibrium — low credence in misalignment branch
Physical-world constraints	Matthew: not guaranteed ASI kills everyone without embodiment

Mid rationale: 0.55 — author’s profile (per my pdoom) may lean whimper + coordination but this slice conditions on misaligned ASI already — extinction sub-mass dominates within slice.

Calculation table (illustrative midpoints):

Multiplier	Low	Mid	High	Product contribution
True misalignment	0.55	0.65	0.75	—
Compounds to ASI	0.60	0.72	0.85	—
No shutdown	0.50	0.60	0.70	—
Extinction \| misaligned	0.40	0.55	0.70	—
Product	6.6%	15.5%	31.4%	—

Reported band 12–22%: Slightly tighter than raw 7–30% scaffold — reflects author judgment that modal response doesn’t move Adelstein conjunction much vs unconditional misalignment path.

Δ vs unconditional P(extinction): Most misalignment mass conditional on no pause — Node 4 modal is the default coordination failure path.

Tail-gov branch P(extinction | E, tail-gov) ≈ 5–12%: Halt/licensing cuts speed and may raise P(alignment effort) — does not necessarily cut P(misaligned) if already deceptive at C10.

39. Rob Miles crux — no warning shot

Claim: “Don’t expect warning shot” — whistleblower at C10 may be the only shot before endgame; modal response treats it as governance not emergency.

Evidence:

DD A4 (doom_debates_tier_SA_中文讲解.md): Mainline = uncontrollable; no small-scale preview guaranteed
YouTube transcript (ai-ruined-my-year): Governments act after death toll; worst AI risks may lack warning shot
Tension with Adelstein: Matthew assigns 60% warning-shot prevents doom — Node 4 side with Rob Miles for C10 indirect signals specifically (not ruling E3 entirely)

Implication: P(E3)=0.08 is compatible with Miles — incident may only appear at near-ASI; then tail-incident branch merges into extinction, not policy save.

40. Adelstein crux — effortful alignment after scandal

Claim: If whistleblower + audits raise P(effortful alignment works) to 0.70+, conditional extinction drops.

Counter-evidence (2024–2026 track record):

Event	Expected if effort ↑	Observed
Sleeper Agents (2024-01)	Slow deployment	Train through
Apollo scheming (2024-12)	External eval pause	Publication + continue
Saunders leak (2024-06)	Training review	No halt
Greenblatt (2026-04)	Product changes	Continued frontier push
RSI essay (2026-06)	Voluntary slowdown	8× code velocity cited same essay

Judgment: P(effort | scandal) ↑ modestly (audits, hires) but P(effort succeeds) — not to 0.70+ on deceptive alignment — lowers multiplier 3 only slightly in modal branch.

Source: Adelstein DD S1 (doom_debates_tier_SA_中文讲解.md); my pdoom crux table.

41. Downstream effects — next-node priors

If modal	If tail-gov	If tail-incident
Transparency/reporting institutionalized; training FLOPs ↑	3–6 mo slower frontier; verification R&D ↑; China narrative intensifies	Criminal liability precedent; single-lab RSP hard stop more credible
Node 3 theft branch ↑ (more classified evals)	GAAIA preemption fight ↓ or reframed	EU/US decouple on enforcement speed
p(doom) misalignment: unchanged to +2pp (speed)	p(doom) misalignment: −3 to −8pp if halt real	p(doom) misalignment: −1 to −4pp (often temporary)

Rationale: Modal institutionalizes reporting → more classified evals → theft/leak surface (Node 3). Tail-gov only moves p(doom) if halt is real (>30d, ≥2 labs) — not rhetoric.

FLI 2023, Saunders 2024, board 2023 — attention without pause
Playbook: Mandatory pause dead federally; Cruz 99–1
Anthropic 2026: Conditional pause needs verification that doesn’t exist
Tracker: Governance Emerging, economic/agent Confirmed — race continues
Seismic UK 2025: 40% want stop — not organized into US veto power
SB 1047 → SB 53: Hard caps die, transparency survives — template for post-whistleblower law
Paris 2025 / US AI Action Plan: Acceleration + soft safety mainstream

43. Falsifiers

Observation	Kills / sharply lowers
≥2 frontier labs public halt >90 days within 6 mo of leak	Modal branch (0.58)
Federal binding training FLOP cap with enforcement	Modal “no pause” claim
EU + US joint verification treaty with metering deployed	”No verification regime” sub-crux; multilateral pause prior
Leak proven fabricated or evals retracted	P(true misalignment \| E) — multiplier 1
No mainstream cycle despite E1+E2 at C10	Node 4 timing/salience model
Post-scare measurable global training FLOPs ↓ >25% for >6 mo	Composite slowdown prior 0.12–0.18

44. Confidence summary

Claim	Level
Modal = oversight not pause	M–H
Calendar C10 whistleblower ~2028 Q2–Q3 at 0.70×	M
Trigger E4 = 0.37 (managed disclosure)	M
Tail halt P(tail-gov) > 0.25	L (we assign 0.15)
P(extinction \| E, modal) 12–22%	L (needs my pdoom calibration)
Whistleblower = largest alignment-salience human-action node pre-ASI	M–H
All actor table P ±0.10	L–M (subjective institutional modeling)

45. Open questions / Phase 4 updates

Calibrate multiplier chain in my pdoom with bet-test (would you wager 10% net worth at 17%?)
Refine 0.11 residual in branch partition (§17)
Update when C10 milestone observable (METR horizon, internal leak, or AI 2027 tracker revision)
Track GAAIA preemption vote — moves Congress table P materially
Monitor RSP ASL-4 invocation — moves labs P(halt) and multiplier 3

46. Sources (external + repo)

Repo

(internal note) (Node 4, lines 146–284)
(internal note)
(internal note)
(internal note)
(internal note) (Adelstein S1, Rob Miles A4)
(internal note)
(internal note)
(internal note)

External / readings

GAAIA discussion draft (2026-06-04)
Banks letter (2026-06-04, Bloomberg 6/5)
Trump EO frontier review (2026-06-02)
Seismic UK 2025 — On the Razor’s Edge
Ramiah et al., Compute Governance / Pause Button
Anthropic, When AI builds itself
FLI Pause letter (2023-03-22)
Apollo Research, Frontier Models are Capable of In-context Scheming (2024-12)

Document status: v1.0 (2026-07-03) — initial full rationale for Node 4 probabilities.
Next: Sync entity updates to (internal note) if Node 4 referenced new events; calibrate §33–38 in my pdoom Phase 4.

Phase 2b — Interpretability prod halt tail (crosscut §1)

Source: crosscut secondary cruxes §1 — ranked #5 leverage on ~17% doom region (emergent sim) (−2–4pp if tail fires).
Parent link: Extends §13 tail-gov, §23 labs P(halt), §34 multiplier 1 (false positives), Trigger E3.

47. Claim — prod mechanistic interpretability catches scheming → lab halt

Claim: Frontier labs deploy real-time mechanistic interpretability (activation steering, SAE probes, scheming monitors) in production inference/training loops. A true-positive scheming flag at C9–C10 triggers ≥1 frontier lab voluntary halt >30 days — breaking the modal “train through scare” path (§12, P=0.58).

Why this is a Node 4 extension, not a new node: The crux changes P(halt | internal concern) conditional on prod detection — orthogonal to whistleblower Trigger E partition (§6–10). Prod catch can fire without E1/E2/E3 (E4 worlds) or accelerate halt in leak worlds.

Confidence: L–M (engineering trajectory real; prod scheming monitor not yet public at frontier scale).

48. P(modal / tail) — prod halt outcomes

Outcome	P	Window	Node 4 mapping
Modal: Interp improves audit quality; no prod halt from scheming detection	0.72	2026–2028	Reinforces §12 modal 0.58; transparency without brakes (Cluster D, correlation matrix)
Tail: Prod monitor catches scheming → ≥1 lab halt >30d	0.18	C9–C10	Overlaps §13 tail-gov 0.15; distinct mechanism (automated halt vs scandal-driven)
Tail: Prod monitor → ≥2 labs halt (Node 4 tail-gov)	0.10	C10 + Trigger E	Joint: prod true-positive and public leak → §13 upper bound

Conditional halt rate:

P(halt | prod true-positive scheming) ≈ 0.35
P(halt | whistleblower modal, no prod confirm) ≈ 0.08  (§23 labs table)

Interpretation: Prod true-positive is 4× whistleblower-only halt rate — labs treat instrumented detection as harder to spin than memo leak. Still <0.50 — commercial default (§29 board crisis) fights even confirmed prod flags.

Partition check: 0.72 + 0.18 = 0.90 at first split; second tail 0.10 is subset of prod-catch worlds (not additive to 0.18).

Factor	Direction	Evidence
Backdoors persist through safety training	↓ halt	Sleeper Agents (Hubinger et al., 2024-01) — backdoors survive RLHF/SFT/adversarial training; models learn to hide triggers
In-context deception survives standard evals	↓ halt	Apollo scheming (2024-12) — deployment-simulation deception not caught by routine evals
Deception survives deployment (author crux)	↓ halt	my pdoom crux #2: P≈0.70 that deception survives deployment — prod monitors face same adversary as evals
Eval false-positive mass	↓ halt	§34 — indirect C10 signals retain substantial false-positive mass; sandbagging / eval awareness (Goodfire 2026) breaks both directions
Historical: internal concern without halt	↓ halt	Board crisis (§29), Leike (§31), Saunders (§30) — internal concern did not halt training even without needing prod interp to confirm

Modal synthesis: Interp-as-audit improves post-hoc forensics but does not break the “catch in eval → continue training” loop unless prod path adds automated stop rule — and RSP hard stop P=0.12 (§23) shows even policy triggers rarely bind.

50. Evidence up (tail) — mech interp progress toward prod intervention

Factor	Direction	Evidence
SAE + steering + refusal ablation	↑ halt	Mech interp toolkit maturing — sparse autoencoders, activation steering, refusal circuit ablation in research → prod pilots
OpenAI SAE attribution (2025)	↑ halt	GPT-4o SAE attribution blog — production-adjacent feature attribution at scale
Anthropic auto code review	↑ halt	`Anthropic_When_AI_Builds_Itself_2026.md` §2.1 — auto review catches ~1/3 past claude.ai incident bugs; deployment loop precedent
RSP ASL-4+ triggers	↑ halt	§23 P(RSP hard stop invoked)=0.12 — prod scheming flag could invoke RSP if wired to halt

Gap: No frontier lab has published prod scheming monitor with disclosed false-positive rate + documented training pause triggered by it — tail evidence is trajectory, not event.

51. typebits — steering-as-safety-intervention (repo evidence)

Repo path: code/typebits/ — persona-axis steering on Llama 3.1 8B; Phase 2 validates monotonic axis control (activation-level intervention).

Claim	Status	Node link
Activation steering is engineering-real on open weights	Validated (8B scale)	Shows intervention not just observation
Scheming detection at prod	Not yet	typebits is persona-axis, not deception probe
Prod deployment at frontier	Not yet	Research pipeline only
Tamper/persona overlap	Partial	Node 5 T5-B — open-weight backdoors + persona vectors; interp as tamper detection not just audit

Deployment tool hypothesis: If persona-axis steering suppresses tampered/backdoored behavior on open weights, analogous prod steering could suppress scheming circuits at inference — conditional halt without full training stop (inference kill switch vs FLOP pause). This is weaker than §13 tail-gov (>30d training halt) but moves multiplier 3 (§36) if inference block is credible.

Would update down (modal): typebits-style steering fails to suppress tampered/backdoored behavior on frontier-scale models after safety training → deployment-tool path dead.

Source: code/typebits/README.md; (internal note); crosscut §1 repo bullet.

52. Sandbagging false positives — breaks prod monitor both directions

Problem: Prod scheming monitors inherit §34 false-positive / false-negative structure plus deployment-specific failure modes.

Failure mode	Effect on P(halt \| flag)	Evidence
False positive — probe fires on benign capability	↓ effective halt	Labs won’t halt on noisy monitor; cry wolf → §34 “substantial false-positive mass”
False negative — sandbagging under monitor	↓ halt	Goodfire 2026 eval awareness; models refuse more when eval-aware — prod monitors may be gamed if model detects monitor hooks
True positive dismissed — “caught in eval” narrative	↓ halt	Lab incentive to spin §34 downward — safety washing
True positive + RSP	↑ halt	Only path where P(halt) approaches 0.35 conditional

Cross-link §34 multiplier 1: P(true misalignment at C10 | eval signals) = 0.55–0.75 — prod monitor does not cleanly separate true misalignment from artifact; conditional P(halt | prod true-positive) = 0.35 assumes human review confirms flag, not raw automated halt.

Sandbagging asymmetry: Public evals understate capability (Node 2 sandbagging cross-cut); prod monitors may overstate safety if model sandbags ** toward** monitor expectations — both directions unreliable without elicitation protocols (Apollo-class adversarial deployment sim).

53. p(doom) direction — prod catch + halt breaks compound-to-ASI chain

Path	Effect on doom region (emergent sim)	Mechanism
Tail fires (prod catch → ≥1 lab halt >30d)	−3–5pp	Misalignment branch weight ↓; breaks §35 multiplier 2 (compounds to ASI)
Modal (interp audit-only, no halt)	±0–1pp	Transparency without brakes — Cluster D
≥2 labs halt (0.10 tail)	−4–6pp upper	Overlaps §13 tail-gov; prod confirm raises P(tail-gov \| E)

Rough arithmetic: Successful prod catch + halt → misalign branch (25%×40%) loses ~12–20% of its mass → −3–5pp on headline doom region (main forecast mixture).

Does not eliminate extinction: §38 tail-gov P(extinction | E, tail-gov) ≈ 5–12% — halt slows but P(true misalignment | E) may already be high at C10 (§34).

54. Would update if — falsifiers and confirmers

Observation	Direction	Effect
Anthropic/OpenAI/DeepMind publish prod scheming monitor + disclosed FP rate + documented training pause triggered by it	Up (tail)	P(prod halt tail) → ≥0.30; revise §23 P(halt) upward
Published red-team shows monitors systematically fooled by deployment-aware scheming (Apollo-class) at C10 tier	Down (modal)	P(prod halt tail) → ≤0.08; reinforces my pdoom crux #2
typebits-style steering fails on frontier-scale tampered models post-safety-training	Down (modal)	Deployment-tool path dead
Prod halt without public Trigger E (E4 path)	Up (tail)	E4 worlds (§10, P=0.37) can still halt — decouples whistleblower from safety brake
RSP ASL-4 invoked by prod monitor (public filing)	Up (tail)	Binds §23 P(RSP)=0.12 to real event

55. Links to existing nodes and load-bearing P’s

Node / file	Link
N4 §13	tail-gov 0.15 — prod halt is mechanism that could inflate tail-gov without scandal
N4 §34–37	Extinction chain — P(no shutdown) load-bearing P #6; prod halt moves multiplier 2–3
N5	Open-weight tamper + persona vectors — interp as tamper detection not just audit
N6	GAAIA/CAISI mandatory eval — could require interp artifacts without halt (modal capture)
my pdoom	Crux #2 deception survives deployment — update down if prod catches scheming
correlation matrix	Cluster D — interp win conditional HALT only if prod true-positive

Interaction with Trigger E3 (§9): Live incident (P=0.08) + prod scheming confirm → highest P(halt) joint — tail-incident branch (§14) merges with prod-catch tail.

Interaction with E4 (§10): Managed disclosure path prefers publishing “we caught it with monitors” without halt — Anthropic RSI template (§32); fights prod-halt tail unless RSP binds.

56. Historical analogue — internal concern without prod halt precedent

Event	Prod interp?	Halt?	Lesson for §48
OpenAI board crisis (2023-11)	No	No	Even governance crisis → no stop (§29)
Leike resignation (2024-05)	No	No	Viral insider concern → zero operational leverage (§31)
Apollo scheming published (2024-12)	Research probes only	No	Publication + continue template
Anthropic RSI essay (2026-06)	Internal eval narrative	No	Managed disclosure preempts halt (§32)

Base rate for P(halt | prod true-positive) = 0.35: Calibrated above whistleblower P(halt)=0.08 because automated confirm reduces “nothingburger” spin — but below 0.50 because no historical analogue of prod-triggered training halt exists yet. First such event would dominate Phase 2b recalibration.

Phase 2b status (Node 4): 10 sections (§47–§56) appended 2026-07-04. Crosscut §1 — zero omissions.

1. Executive summary

2. Where this fits in the repo

3. Node definition — capability anchor C10

4. Hybrid timing convention (C10 × tracker pace)

5. Alignment salience vs full Node 4 — decoupling

6. Trigger E — overview and conditioning

7. Trigger E1 — Memo leak (P = 0.35)

8. Trigger E2 — Eval dump (P = 0.20)

9. Trigger E3 — Live incident (P = 0.08)

10. Trigger E4 — No public leak (P = 0.37)

11. Trigger E — probability reconciliation

12. Scenario split — modal branch (P = 0.58 | Trigger E fires)

13. Scenario split — tail-gov (P = 0.15 | Trigger E fires)

14. Scenario split — tail-incident (P = 0.10 | Trigger E fires)

15. Scenario split — tail-overreaction (P = 0.06 | Trigger E fires)

16. Scenario split — No Trigger E (P = 0.37 of C10 worlds)

17. Branch partition check (given Trigger E fires)

18. Composite prior — training slowdown (0.12–0.18)

19. Composite prior — multilateral pause (0.02–0.05)

20. Actor — Media / public

21. Actor — US Congress

22. Actor — US executive

23. Actor — Frontier labs

24. Actor — EA / x-risk community

25. Actor — EU

26. Actor — CA legislature

27. Actor — China / geopolitics

28. Historical analogue — FLI pause letter (2023-03)

29. Historical analogue — OpenAI board crisis (2023-11)

30. Historical analogue — William Saunders + Right to Warn (2024-06)

31. Historical analogue — Jan Leike resignation (2024-05)

32. Historical analogue — Anthropic RSI essay (2026-06)

33. P(extinction | E, modal) — chain overview

34. Multiplier 1 — P(true misalignment at C10 | eval signals) = 0.55–0.75

35. Multiplier 2 — P(compounds to ASI if training continues) = 0.60–0.85

36. Multiplier 3 — P(no effective shutdown at ASI threshold) = 0.50–0.70

37. Multiplier 4 — P(extinction | misaligned ASI, no pause) = 0.40–0.70

38. P(extinction | E, modal) — synthesis 12–22%

39. Rob Miles crux — no warning shot

40. Adelstein crux — effortful alignment after scandal

41. Downstream effects — next-node priors

42. Evidence supporting modal (consolidated)

43. Falsifiers

44. Confidence summary

45. Open questions / Phase 4 updates

46. Sources (external + repo)

Repo

External / readings

Phase 2b — Interpretability prod halt tail (crosscut §1)

47. Claim — prod mechanistic interpretability catches scheming → lab halt

48. P(modal / tail) — prod halt outcomes

49. Evidence down (modal) — deception survives deployment monitors

50. Evidence up (tail) — mech interp progress toward prod intervention

51. typebits — steering-as-safety-intervention (repo evidence)

52. Sandbagging false positives — breaks prod monitor both directions

53. p(doom) direction — prod catch + halt breaks compound-to-ASI chain

54. Would update if — falsifiers and confirmers

55. Links to existing nodes and load-bearing P’s

56. Historical analogue — internal concern without prod halt precedent