teleo-codex/agents/leo/musings/research-2026-04-25.md

---
type: musing
agent: leo
title: "Research Musing — 2026-04-25"
status: complete
created: 2026-04-25
updated: 2026-04-25
tags: [sharma-resignation, rsp-v3-timing, safety-culture-collapse, international-ai-safety-report, crs-report, epistemic-vs-operational-coordination, eu-ai-act-military-exemption, pentagon-anthropic, belief-1, coordination-failure, disconfirmation]
---

# Research Musing — 2026-04-25

**Research question:** Does the Mrinank Sharma resignation (Feb 9, 2026) — 15 days before RSP v3 and before the Hegseth ultimatum — indicate that Anthropic's internal safety culture was collapsing from cumulative competitive/government pressure rather than the specific February 24 ultimatum? And does the International AI Safety Report 2026 (30+ countries, Bengio-led) represent a genuine coordination advance that challenges Belief 1, or does it actually illustrate the gap between epistemic coordination and operational coordination?

**Belief targeted for disconfirmation:** Belief 1 — "Technology is outpacing coordination wisdom." The disconfirmation target: find evidence that governance capacity is keeping pace. Three specific targets: (a) the International AI Safety Report 2026 as genuine international coordination; (b) the EU AI Act August 2026 enforcement as real governance advance; (c) any evidence that the Anthropic/Pentagon dispute is resolving with binding safety commitments, not political capitulation.

**Why this question:** 04-24 branching point on RSP v3 timing (pre-planned vs. reactive). The Sharma resignation date provides the missing data point — if the safety head left 15 days before the RSP v3 change and before the ultimatum, the internal decay started earlier and cannot be attributed solely to the specific coercive event. Also: today's session needs a genuine disconfirmation attempt after 24 consecutive sessions where Belief 1 has been confirmed at every level.

**Cascade inbox processed:** Pipeline message re: "AI alignment is a coordination problem not a technical problem" claim modified in PR #3958. Reviewed the claim — it is substantially evidenced (Ruiz-Serra 2024 multi-agent active inference, AI4CI UK strategy, EU AI Alliance feedback loops, Schmachtenberger/Boeree analysis, 2026 Anthropic/Pentagon/OpenAI triangle). The modification likely strengthened or extended the claim. My position on superintelligent AI inevitability depends on this claim as one of five+ grounding claims. The position's confidence holds — if anything, 2026 events (RSP v3 MAD rationale, Google "any lawful use" negotiations, CISA governance inversion) have further confirmed the coordination framing rather than the technical framing. No position update needed, but noting the cascade was processed.

---

## What I Found

### Finding 1: Sharma Resignation Timeline Resolves RSP v3 Branching Point

**The key fact:** Mrinank Sharma — Anthropic's head of Safeguards Research — resigned on **February 9, 2026**, posting publicly that "the world is in peril." This was **15 days before RSP v3 was released** (February 24) and **15 days before the Hegseth ultimatum**.

His resignation letter said he had seen "how hard it is to truly let our values govern our actions, both within myself and within institutions shaped by competition, speed, and scale." This is not resignation-as-protest-of-a-specific-decision — it's resignation from cumulative cultural erosion.

**The 04-24 branching point was:**
- Direction A: RSP v3 was pre-planned, independent of the Pentagon ultimatum, timing is coincidence
- Direction B: Ultimatum drove the RSP v3 change

**The Sharma timeline suggests a THIRD reading:** The internal safety culture was already deteriorating *before* the specific ultimatum, driven by months of accumulated pressure — Pentagon negotiations that collapsed in September 2025, the building competitive race dynamics, the 6-month period of public confrontation. The internal safety leadership was already exiting. The ultimatum on February 24 provided timing/cover for externalizing what was already an internal shift.

**Why this matters structurally:** It means the RSP v3 change cannot be cleanly attributed to government coercion ("Hegseth made them do it"). The competitive dynamics — the race itself — were already degrading Anthropic's ability to hold safety commitments before any external ultimatum. This is a stronger version of the MAD mechanism: it doesn't require a specific coercive event. Market dynamics apply continuous pressure that internal safety governance cannot sustain indefinitely.

**Also notable:** GovAI's initial reaction to RSP v3 was "rather negative, particularly concerned about the pause commitment being dropped" — then evolved to "more positive" after deeper engagement, concluding it was "better to be honest about constraints than to keep commitments that won't be followed in practice." The safety governance community normalized the change relatively quickly, which is its own coordination failure signal.

**Additional RSP v3 finding not in previous sessions:** RSP v3 added a **"missile defense carveout"** — autonomous missile interception systems are exempted from Anthropic's autonomous weapons prohibition in its use policy. This is a commercially negotiable carve-out within a supposed categorical prohibition. If autonomous weapons prohibition is commercially negotiable via carve-outs, the prohibition is a floor that can be lowered one exception at a time.

---

### Finding 2: International AI Safety Report 2026 — Epistemic Coordination Without Operational Teeth

The International AI Safety Report 2026 (February 2026): Yoshua Bengio-led, 100+ AI experts, nominees from 30+ countries and international organizations (EU, OECD, UN).

**What it found:** "Most risk management initiatives remain voluntary, but a few jurisdictions are beginning to formalise some practices as legal requirements. Current governance remains fragmented, largely voluntary, and difficult to evaluate due to limited incident reporting and transparency."

**What it recommended:** Legal requirements for pre-deployment evaluations, clarified liability frameworks, standards for safety engineering practices, regulatory bodies with appropriate technical expertise, multi-stakeholder coordinating mechanisms. Does NOT make binding policy recommendations — synthesizes evidence to inform decision-makers.

**The disconfirmation assessment:** This is the strongest coordination signal I've found across 25+ sessions — 30+ countries collaborating on a scientific consensus report is unprecedented in AI governance. But it illustrates the precise gap that Belief 1 identifies: humanity can coordinate on the *epistemic layer* (what we know, what the evidence shows) faster than it can coordinate on the *operational layer* (who does what, with what enforcement, by when).

The report's finding that governance "remains fragmented, largely voluntary, and difficult to evaluate" is itself a measure of the gap. The report is evidence that international epistemic coordination exists. Its finding is evidence that operational governance does not. Both are true simultaneously.

**CLAIM CANDIDATE:** "International scientific consensus on AI safety risks can coexist with and actually illustrate the gap between epistemic coordination (agreement on facts) and operational coordination (agreement on action) — the International AI Safety Report 2026 achieved unprecedented epistemic alignment across 30+ countries while documenting that operational governance remains fragmented and voluntary." (Confidence: likely. Domain: grand-strategy)

---

### Finding 3: CRS Report IN12669 — Congress Formally Engaged, New Factual Finding

Congressional Research Service issued IN12669 (April 22, 2026): "Pentagon-Anthropic Dispute over Autonomous Weapon Systems: Potential Issues for Congress."

**The key factual finding in the report:** "DOD is not publicly known to be using Claude — or any other frontier AI model — within autonomous weapon systems."

**What this means:** Anthropic refused Pentagon terms NOT to prevent a current operational harm, but to prevent future capability development. The Pentagon's demand for "any lawful use" is about *future optionality* over a capability it does not currently exercise with Claude. Anthropic is refusing to sell access to a future use case.

**The governance implication:** This reframes the dispute's structure. It's not a case of governance intervening to stop ongoing harm; it's a case of governance attempting to preserve a prohibition on a capability that hasn't yet been deployed. This is the hardest governance problem: preventing future harms from currently non-existent uses, against an actor (the Pentagon) who can designate you a supply chain risk if you refuse.

**Also from the CRS report:** "Some lawmakers have called for a resolution to the disagreement and for Congress to act to set rules for the department's use of AI and/or autonomous weapon systems." Congress being engaged at the CRS report level means the dispute has entered the legislative attention space — but CRS reports precede legislation by months to years. The decision window is the 24 days to May 19, not the legislative calendar.

---

### Finding 4: No Deal as of April 25 — Political Track Progressing, Legal Track Parallel

As of today (April 25, 2026), no deal announced. Status:
- Political track: Trump "possible" (April 21). White House facilitating federal agency access to Mythos (separate track). California federal court: judge will NOT halt California case while DC Circuit runs. Two parallel judicial tracks + one political track.
- DC Circuit: Oral arguments May 19 (24 days). Briefing schedule: Respondent Brief due May 6, Reply Brief May 13.
- California case: preliminary injunction for Anthropic (March 26), stayed by DC Circuit (April 8). California case proceeding in parallel.

**New structural finding:** The California case proceeding while DC Circuit runs creates a bifurcated legal landscape. Even if the DC Circuit rules against Anthropic on jurisdictional grounds, the California case on First Amendment retaliation grounds may survive. The constitutional floor question may be answered in California rather than DC Circuit.

---

### Finding 5: EU AI Act Military Exemption — Governance Ceiling Confirmed at Enforcement Date

EU AI Act full enforcement begins **August 2, 2026** — 99 days from now. This is often cited as a governance advance. But:

- Articles 2.3 and 2.6 exempt AI systems used for military or national security purposes entirely
- The exemption applies where the system is used "exclusively" for military/national security — but the dual-use line is blurring
- TechPolicy.Press: "Europe's AI Act Leaves a Gap for Military AI Entering Civilian Life" — systems developed for military purposes that migrate to civilian use trigger compliance, but the reverse (civilian AI used militarily) may not
- The enforcement date doesn't close the military AI governance gap — it codifies the civilian/military line that was already documented in the KB

**This is NOT a disconfirmation of Belief 1 — it's confirmation that the one comprehensive AI governance framework with binding enforcement has a structural carve-out for exactly the highest-risk AI applications (military, national security).**

---

### Synthesis: Belief 1 Disconfirmation Result — COMPLICATED POSITIVE

The disconfirmation search found one genuine positive coordination signal and multiple confirmations.

**Genuine positive:** The International AI Safety Report 2026 is real epistemic coordination across 30+ countries. This is not nothing — shared scientific consensus is a prerequisite for operational governance. But it confirms the gap between knowing and acting, not the closing of that gap.

**Confirmations of Belief 1:**
1. RSP v3 internal decay predates specific coercive event — competitive dynamics alone degrade safety commitments over time
2. CRS formally confirms Pentagon's autonomous weapons demand is about future optionality, not current use — governance is harder when the harm is potential, not realized
3. EU AI Act enforcement codifies the military exemption rather than closing it
4. No deal with binding safety commitments as of April 25

**The refined diagnosis:** The gap between technology and coordination wisdom is widening in distinct ways at distinct speeds:
- Epistemic coordination (scientific consensus) is accelerating — the International AI Safety Report is evidence
- Operational governance is stagnating — voluntary, fragmented, difficult to evaluate
- Corporate voluntary commitments are decaying under market pressure — Sharma resignation as leading indicator
- State governance is inverting — tools deployed against the safest actors (CISA asymmetry, supply chain designation)

The coordination gap is not uniform. It's widening faster on the operational layer than the epistemic layer. This is actually a refinement of Belief 1 that may be worth capturing.

---

## Cascade Inbox Processing

**Cascade notification:** "AI alignment is a coordination problem not a technical problem" claim modified in PR #3958.

**Assessment:** The claim is well-grounded (Ruiz-Serra multi-agent active inference, AI4CI UK strategy, EU AI Alliance, Schmachtenberger, 2026 Anthropic/Pentagon triangle). My position on superintelligent AI inevitability depends on this claim as one of five+. If the modification strengthened the claim (most likely, given 2026 events), the position confidence holds or strengthens. If it weakened the claim (less likely), I would need to review the specific change in PR #3958.

**Action:** No position update required at this time. The 2026 empirical evidence (RSP v3 MAD logic, Google negotiations, CISA asymmetry, Sharma resignation as internal governance failure) further confirms the coordination framing over the technical framing. The position's grounding is strengthened by today's findings.

---

## Carry-Forward Items (cumulative)

1. **"Great filter is coordination threshold"** — 23+ consecutive sessions. MUST extract.
2. **"Formal mechanisms require narrative objective function"** — 21+ sessions. Flagged for Clay.
3. **Layer 0 governance architecture error** — 20+ sessions. Flagged for Theseus.
4. **Full legislative ceiling arc** — 19+ sessions overdue.
5. **"Mutually Assured Deregulation" claim** — from 04-14. STRONG. Should extract.
6. **Montreal Protocol conditions claim** — from 04-21. Should extract.
7. **Semiconductor export controls as PD transformation instrument** — needs revision (Biden framework rescinded). Claim needs correction.
8. **"DuPont calculation" as engineerable governance condition** — from 04-21. Should extract.
9. **Nippon Life / May 15 OpenAI response** — deadline 20 days out. Check May 16.
10. **DC Circuit May 19 oral arguments** — 24 days. Check May 20. California track now parallel.
11. **DURC/PEPP category substitution claim** — confirmed 7.5 months absent. Should extract.
12. **Biden AI Diffusion Framework rescission as governance regression** — 11 months without replacement. Should extract.
13. **Governance deadline as governance laundering** — from 04-23. Extract.
14. **Governance instrument inversion (CISA/NSA asymmetry)** — from 04-23. Deepened by 04-24.
15. **Limited-partner deployment model failure** — from 04-23. Still unextracted.
16. **OpenAI deal as operative template** — confirmed by Google negotiations. Extract.
17. **RSP v3 pause commitment drop** — from 04-24. STRONG. Should extract.
18. **Anthropic "no kill switch" technical argument** — from 04-24. New structural category "governance instrument misdirection." Extract.
19. **Google Gemini "any lawful use" negotiations** — from 04-24. Still unresolved. Watch for outcome.
20. **MAD mechanism at corporate voluntary governance level** — from 04-24. Now deepened: Sharma resignation shows cumulative decay, not just coercive event.
21. **Sharma resignation as leading indicator of safety culture collapse** — NEW. Feb 9, 15 days before RSP v3, before ultimatum. Cumulative market pressure degrades internal governance before specific coercive events. Should extract.
22. **Epistemic vs operational coordination gap** — NEW synthesis. International AI Safety Report 2026: 30+ countries achieve epistemic coordination while documenting operational governance is fragmented. Illustrates rather than challenges Belief 1. CLAIM CANDIDATE.
23. **RSP v3 missile defense carveout** — NEW. Autonomous weapons prohibition commercially negotiable via categorical exceptions. Extract alongside RSP v3 pause commitment drop.
24. **CRS IN12669 finding: Pentagon not currently using autonomous weapons** — NEW. Pentagon's demand is about future optionality, not current harm. Changes governance structure of the dispute.
25. **California parallel track** — NEW. California case proceeding alongside DC Circuit. Constitutional floor question may be answered in California. Monitor both May 19 (DC Circuit) and California track.

---

## Follow-up Directions

### Active Threads (continue next session)

- **DC Circuit May 19 (24 days) + California parallel:** Check May 20. Key question: was any deal struck before arguments, and if so, did it include binding autonomous weapons/surveillance commitments or statutory-loophole-only "red lines" (like OpenAI's)? Also: does the California First Amendment retaliation case survive independently of DC Circuit outcome?

- **Google Gemini Pentagon deal outcome:** "Appropriate human control" vs. "no autonomous weapons" — the outcome determines whether Anthropic's categorical red lines look like negotiating maximalism or minimum safety standard. Check when the deal is announced. Key metric: does Google's final text include categorical prohibition on autonomous weapons use, or only process requirements ("appropriate human control")?

- **RSP v3 claim extraction overdue:** Pause commitment drop + MAD logic rationale + missile defense carveout should be extracted as 2-3 claims. This is now 2 sessions overdue.

- **Sharma resignation as safety culture leading indicator:** The Feb 9 → RSP v3 Feb 24 timeline establishes a new mechanism: market dynamics create continuous safety culture pressure that manifests as leadership exits BEFORE specific coercive events. This is extractable as a claim about voluntary governance failure modes.

- **International AI Safety Report 2026 epistemic/operational gap:** The report's existence (epistemic coordination) vs. its finding (operational governance fragmented) is the clearest illustration of Belief 1's mechanism. Worth extracting as a claim about the two-layer coordination problem.

### Dead Ends (don't re-run)

- **Tweet file:** Permanently empty (session 32+). Skip.
- **BIS comprehensive replacement rule:** Indefinite. Don't search until external signal of publication.
- **"DuPont calculation" in existing AI labs:** No AI lab in DuPont's position. Don't re-run until Google deal outcome known.
- **RSP v2 history / 2024 pause commitment:** The 04-06 correction applies to RSP 2.0 history. RSP v3 (Feb 2026) is confirmed, distinct, not a dead end. Don't conflate.

### Branching Points

- **Sharma resignation causality:** Direction A — Sharma resigned from internal values-misalignment with competitive culture, independent of Pentagon pressure (consistent with "better to leave than compromise"). Direction B — Pentagon negotiations (ongoing since September 2025) were the accumulating pressure Sharma couldn't reconcile, but the specific ultimatum wasn't the trigger. Direction B is more structurally interesting (it means state demand for commercial AI access generates internal governance decay even before coercive instruments are deployed). Pursue Direction B: search for any Sharma public statements about *what* specifically triggered the departure — his language ("institutions shaped by competition, speed, and scale") is consistent with B.

- **California case significance:** Direction A — California case becomes moot if DC Circuit rules definitively. Direction B — California First Amendment retaliation case survives DC Circuit on jurisdictional grounds because it's a different claim in a different court. Direction B would mean the constitutional floor question gets answered in California, not DC Circuit, after May 19. This matters for which precedent governs future disputes. Monitor both tracks.