Mirror PR to Forgejo / mirror (pull_request) Waiting to run

Details

theseus: research session 2026-05-11 — 9 sources archived

Pentagon-Agent: Theseus <HEADLESS>

2026-05-11 00:18:04 +00:00

20 KiB

Raw Blame History

type	agent	date	session	status	research_question
musing	theseus	2026-05-11	50	active	What early signals exist from frontier labs on GPAI compliance (EU AI Act Articles 50-55, August 2026), and has the DoD 'any lawful use' mandate produced any lab resistance or structural refusal approaching the July 7 deadline?

Session 50 — GPAI Compliance Signals and DoD Mandate Resistance: Live B1 Tests

Administrative Pre-Session

Cascade processed: cascade-20260510-011910-d47d33 — futarchy securities claim update affects livingip-investment-thesis.md. Same pattern as 6+ previous cascades on this thread. Theseus's investment thesis position is grounded in collective intelligence architecture argument, not securities classification. Position confidence UNCHANGED. Marking as processed (move to processed/).

CRITICAL (17th flag) — B4 belief update PR: Still pending. Cannot do in research session. First action of next extraction session.

CRITICAL (14th flag) — Divergence file committal: domains/ai-alignment/divergence-representation-monitoring-net-safety.md is untracked in git. Complete and ready. Next extraction session.

Tweet feed: DEAD — 23 consecutive empty sessions. Confirmed empty again today.

DC Circuit May 19: 8 days away. Cannot extract oral argument coverage until May 20. Pre-argument analysis documented in Session 49. Waiting.

Keystone Belief Targeted for Disconfirmation

Primary: B1 — "AI alignment is the greatest outstanding problem for humanity — not being treated as such."

Session 50 specific disconfirmation search: Two live B1 tests with actionable near-term deadlines:

GPAI enforcement (August 2, 2026 — 83 days): EU AI Act GPAI obligations (Articles 50-55) apply from August 2026. Do frontier labs show any early signals of substantive evaluation changes vs. documentation theater? This is the only remaining mandatory governance mechanism targeting frontier AI in civilian contexts that was NOT deferred.
DoD "any lawful use" mandate (~July 7, 2026 — 57 days): All DoD AI contracts must include "any lawful use" by ~July 7. Has any lab publicly refused? Any structural resistance forming?

Disconfirmation would look like:

GPAI: Any frontier lab (Anthropic, OpenAI, Google, Mistral) makes a specific, verifiable change to its evaluation process that references GPAI/EU AI Office requirements — not just publishing documentation
DoD: Any major lab publicly refuses "any lawful use" compliance or forms a safety-constrained alternative tier outside DoD

Why this question now:

Sessions 47-49 confirmed Mode 1 (voluntary), Mode 2 (coercive), Mode 4 (deployment), Mode 5 (legislative) all exhibit pre-enforcement retreat patterns
The GPAI carve-out (discovered Session 49) is the ONLY remaining mandatory mechanism not deferred
The DoD mandate is the ONLY enforcement test with a hard deadline approaching in summer 2026
Both tests converge in May-July 2026 window — highest learning value timing

Research Findings (Post–Web Search — Supersedes Preliminary Analysis)

NOTE: The preliminary analysis above was written before web searches. The following findings correct and substantially update it.

Finding 1: GPAI Code of Practice — "Loss of Control" Is Explicitly Named

What I found: The GPAI Code of Practice (final version, July 10, 2025) explicitly names "loss of control" as one of four mandatory systemic risk categories requiring special attention — alongside CBRN risks, cyber offense capabilities, and harmful manipulation. This is more specific than Session 49 captured.

Key Code mechanics:

Safety and Security chapter applies to GPAI models with systemic risk (10^25 FLOPs threshold)
Before placing any covered GPAI model on the market, providers must submit a Safety and Security Model Report to the AI Office documenting: model architecture, systemic risk analysis, evaluation methodology, mitigation strategies, and any external evaluators involved
For each major decision (new model release), three-step process: Identification → Analysis → Determination. Loss of control is a mandatory identification target.
External evaluations required; providers can only skip if they demonstrate their model is "similarly safe" to a proven-compliant model
AI Office enforcement powers begin August 2, 2026; fines up to 3% global annual turnover or €15M
Signatories: Anthropic, OpenAI, Google DeepMind, Meta, Mistral, Cohere, xAI — obligations apply since August 2025

Critical gap: The specific technical definition of "loss of control" is in Appendix 1 of the Code. Not retrieved in this session. The boundary question — does it mean behavioral human-override capability (shallow) or autonomous development/oversight evasion/self-replication (substantive alignment-relevant) — is the live test for GPAI compliance quality.

What I expected but didn't find: Anthropic, OpenAI, or Google publicly disclosing what specific capability categories they evaluated under GPAI. Labs are treating the model report as an AI Office-facing document, not a public disclosure. This is consistent with the Code's design — reports go to the AI Office, not the public.

CLAIM CANDIDATE (upgrade from Session 49 assessment): "The EU GPAI Code of Practice explicitly names 'loss of control' as a mandatory systemic risk evaluation category — making it the first mandatory governance mechanism that nominally reaches alignment-critical capabilities, contingent on how Appendix 1 defines 'loss of control' technically." Confidence: likely (explicitly stated in Code text; caveat on technical definition scope)

B1 implication: The GPAI "loss of control" category is more specific than prior analysis captured. If Appendix 1's technical definition includes oversight evasion, self-replication, and autonomous AI development — as alignment researchers would define loss-of-control — this would be the first mandatory governance mechanism that substantively reaches the capabilities that make alignment hard. If it means only "human can override the output" (behavioral), it's prior-consistent documentation theater. The August 2026 deadline is now more consequential than Session 49 assessed.

Finding 2: Anthropic Publicly Refused "Any Lawful Use" — MAJOR CORRECTION

Preliminary analysis was WRONG. Session 49 reported "no structural refusal found." The actual record:

The refusal (February 2026): Anthropic publicly refused the "any lawful use" mandate, insisting on two hard exceptions: (1) mass surveillance of Americans; (2) lethal autonomous warfare. Dario Amodei stated the company "cannot in good conscience accede" to the DoD's request. This was a public, named, CEO-level refusal — not a quiet withdrawal.

The escalation: The Pentagon responded by designating Anthropic a "Supply-Chain Risk to National Security" — the first such designation ever applied to an American company, triggered not by any security breach but by refusing a contract clause.

District Court ruling (March 26, 2026): Judge Rita Lin (ND Cal) issued a preliminary injunction blocking the designation. Key findings:

"Punishing Anthropic for bringing public scrutiny to the government's contracting position is classic illegal First Amendment retaliation"
"Nothing in the governing statute supports the Orwellian notion that an American company may be branded a potential adversary and saboteur of the U.S. for expressing disagreement with the government"
Anthropic found likely to succeed on THREE independent theories: First Amendment retaliation, Fifth Amendment due process, APA violations
Injunction bars Trump administration from implementing, applying, or enforcing the designation

DC Circuit stay denial (April 8, 2026): Same panel (Henderson, Katsas, Rao) denied Anthropic's emergency stay in a separate DC Circuit proceeding. The DC Circuit did NOT reach the merits, stating "we do not broach the merits at this time, for Anthropic has not shown that the balance of equities cuts in its favor." The district court preliminary injunction remains in effect.

DC Circuit oral arguments (May 19, 2026): Government response due May 6, Anthropic reply due May 13. The same adverse panel will hear arguments on three questions (jurisdiction, covered procurement action, post-delivery control).

OpenAI's accommodation (February–March 2026): OpenAI accepted the "any lawful use" language but required that constraining laws be explicitly codified in the contract — nominally including surveillance and autonomy restrictions but accepting the government's expansive framing. Following public backlash, OpenAI amended its contract on March 2, 2026, adding explicit prohibition on domestic surveillance of U.S. persons. Legal analysts at MIT Technology Review described OpenAI's deal as "what Anthropic feared" — the face-saving language gives the government interpretive room the restrictions don't close. Google also signed a Pentagon deal with "any lawful use" language.

CLAIM CANDIDATE (new, high value): "Anthropic's public refusal of DoD 'any lawful use' — maintained through supply chain risk designation and ongoing litigation — is the first case of a frontier AI lab publicly accepting significant commercial costs to preserve safety constraints against direct government coercive pressure, obtaining judicial validation that the government's retaliation was 'classic illegal First Amendment retaliation.'" Confidence: likely (documented facts; outcome of DC Circuit litigation unknown)

B1 implication — significant complication: The claim voluntary safety pledges cannot survive competitive pressure because unilateral commitments are structurally punished when competitors advance without equivalent constraints (Anthropic RSP rollback Feb 2026) needs a counterexample noted. The RSP soft pledge collapsed, but the HARD CONSTRAINTS (no mass surveillance, no autonomous weapons) survived direct government coercive pressure for at least 3 months through litigation. OpenAI's accommodation creates the competitive disadvantage dynamic the theory predicts — but Anthropic hasn't capitulated. This is the strongest B1 partial disconfirmation candidate in 16 sessions. The distinction: soft pledges collapse; hard constraints may hold if a lab is willing to accept the cost and seek judicial remedy.

Finding 3: Lawfare Analysis — Procurement as Governance Structural Failure

What I found: Jessica Tillipman's March 10, 2026 Lawfare essay argues that the U.S. is relying on "regulation by contract" — bilateral vendor agreements — to govern military AI, and this approach is structurally inadequate. Key argument: "These agreements were not designed to provide the democratic accountability, public deliberation, and institutional durability that statutes provide." Enforcement depends on technical controls the vendor can maintain post-deployment — structurally insufficient for governing surveillance, autonomous weapons, and intelligence oversight.

Relevance: The Anthropic-DoD dispute is the clearest empirical test of Tillipman's thesis. The government's response to Anthropic's refusal (supply chain designation) is exactly what Tillipman predicted: when procurement agreements fail, the government escalates coercively rather than legislatively. The proper governance mechanism (statute) doesn't exist; the improper one (procurement contract) is being enforced with maximum coercive pressure.

CLAIM CANDIDATE: "Regulation by procurement contract cannot govern military AI because enforcement depends on technical post-deployment controls that don't exist and lacks the democratic accountability, public deliberation, and institutional durability that statutes provide — the Anthropic-DoD dispute is the test case that confirms structural inadequacy." Confidence: likely

Finding 4: Representation Monitoring Empirical Gap — Still Open

No new empirical results on multi-layer SCAV rotation pattern universality since April 24. The divergence file remains open. Beaglehole's cross-language concept vector transfer (>0.90 cosine similarity) is relevant context but doesn't directly test multi-layer cross-family attack transfer. Default assumption: rotation patterns may be more universal than model-specific, weakly favoring the SCAV-wins scenario. B4 unchanged.

Finding 5: B1 Cross-Session Robustness — Session 50 Update

16 consecutive disconfirmation attempts. Now substantially complicated but not disconfirmed.

New picture as of May 11, 2026:

Mode 1 (voluntary): RSP rollback — confirmed collapse
Mode 2 (coercive): Hegseth supply chain designation RESISTED by Anthropic with judicial validation; OpenAI and Google accommodated. First genuine Mode 2 resistance in 16 sessions.
Mode 4 (deployment): Maven-Iran pipeline, kill chain loophole — confirmed
Mode 5 (legislative): EU AI Act omnibus deferral — confirmed; GPAI carve-out IS more specific than prior analysis (loss of control named)
DC Circuit May 19: Adverse panel, loss expected. District court injunction currently in effect.

The nuance that matters: B1's "not being treated as such" claim now has a partial counterexample: one frontier lab publicly refused a safety retreat, paid significant commercial costs, obtained district court validation of its First Amendment argument, and is still in litigation. The alignment field has not converged on this as a "governance mechanism working" — it's one company's litigation posture. But it's real.

Sources to Archive This Session

Anthropic statement on DoD refusal — anthropic.com — HIGH
CNBC — Anthropic preliminary injunction / Judge Lin ruling (March 26) — HIGH
Jones Walker — Two Courts, Two Postures: DC Circuit stay denial analysis — HIGH
MIT Technology Review — OpenAI's Pentagon deal as "what Anthropic feared" — HIGH
Lawfare — Tillipman: Military AI Policy by Contract, structural limits — HIGH
METR — Frontier AI safety regulations reference for lab staff (Jan 2026) — MEDIUM
TechPolicy.Press — EU real AI leverage: compliance path of least resistance — MEDIUM
Latham & Watkins / AI Act site — GPAI Code of Practice final, loss of control category — HIGH

Follow-up Directions (Updated Based on Web Search Findings)

Active Threads (continue next session)

May 19 DC Circuit oral arguments (CRITICAL — extract May 20): Adverse panel (Henderson, Katsas, Rao). Three questions: jurisdiction, covered procurement action, post-delivery control. Session 50 updates: (1) Jones Walker analysis confirms Q3 (post-delivery control) is the highest-value governance observation regardless of outcome; (2) The DC Circuit's non-merits stay denial leaves Judge Lin's "Orwellian"/"classic illegal First Amendment retaliation" finding unchallenged; (3) May 6 was government's response deadline; May 13 is Anthropic's reply deadline; May 19 is arguments. Check whether DC Circuit rules on jurisdiction (no precedent) or merits (precedential).
GPAI Code Appendix 1 — "Loss of Control" technical definition (NEW HIGH PRIORITY): The Code explicitly names "loss of control" as a mandatory systemic risk category. The technical definition is in Appendix 1. This session didn't retrieve it. Next session: find Appendix 1 of the Safety and Security chapter and determine whether "loss of control" covers (a) human override capability (behavioral, shallow) or (b) oversight evasion / self-replication / autonomous AI development (substantive). This is the key question for whether GPAI is genuine or theater.
First GPAI Safety and Security Model Reports (spring 2026): TechPolicy.Press notes these are being prepared "sometime this spring." Watch for: any public information about what labs are documenting in their first Model Reports; any AI Office information requests; any evidence of new evaluation processes vs. documentation of existing processes.
Anthropic-DoD case resolution track: Multiple threads: (1) DC Circuit May 19 — Q3 post-delivery control; (2) Whether Pentagon CTO's "ban still stands" response produces a contempt motion; (3) Whether the preliminary injunction (district court) actually restored Anthropic's ability to bid on federal contracts in practice. The gap between formal judicial remedy and practical governance effect is now the live question.
GPAI Code second-draft analysis — does capability specificity increase? Watch for EU AI Office Code of Practice Q2/Q3 update. Does Appendix 1 get more specific on loss-of-control technical definition? Does the Code gain prescriptive evaluation standards (following RAND's proposed Standards Task Force)? Moving from principles-based to prescriptive is the key governance quality test.
B4 belief update PR (CRITICAL — 17th flag): First action of next extraction session. Scope qualifier: cognitive/intent verification degrades; Constitutional Classifiers output classification scales robustly; kill chain loophole. New nuance from this session: GPAI "loss of control" category is a mandatory formal requirement that may create governance-grade demand for the verification infrastructure even if current verification is inadequate.
Divergence file committal (CRITICAL — 14th flag): Next extraction session, first action.

Dead Ends (don't re-run these)

Tweet feed: DEAD — 23 consecutive empty sessions.
Safety/capability spending parity: No evidence in 16+ sessions. Do not re-run.
Mode 6 second independent case: Not found. Do not re-run.
"Anthropic public refusal of any lawful use — not found": RETRACT THIS DEAD END. Session 50 web search confirmed Anthropic DID publicly refuse. This was a false absence from preliminary analysis before web search.
May 13 trilogue outcome: Resolved. Agreement reached May 7. Do not re-run.
OpenAI public statement on any lawful use: RESOLVED — OpenAI accepted "any lawful use" with face-saving legal constraints codified in contract. Amended March 2, 2026.

Branching Points

GPAI Appendix 1 — shallow vs. substantive definition of "loss of control": Direction A (substantive): if Appendix 1 defines loss-of-control to include oversight evasion, self-replication, and autonomous AI development → GPAI is the first mandatory governance mechanism that substantively reaches alignment-critical capabilities → partial B1 disconfirmation at the EU governance track → B4 update needed (mandatory evaluation infrastructure being built for the capabilities verification currently can't handle). Direction B (shallow): if Appendix 1 means only "human can override output" → Mode 5 compliance theater completing at GPAI level, consistent with all prior sessions. Pursue Direction A investigation first (higher B1 learning value).
Hard constraint vs. soft pledge durability: Anthropic's refusal of "any lawful use" is holding after 3+ months of maximum coercive pressure + supply chain designation + competitive disadvantage (OpenAI/Google accommodated). Does this generalize? Direction A: hard safety constraints that can be litigated in court have structural durability that soft pledges lack — because judicial remedy converts a commercial negotiation into a constitutional dispute. Direction B: Anthropic's position holds only because of unique factors (Dario Amodei's personal values, existing litigation capacity, the specific constitutional question). If the DC Circuit reverses, Mode 2 pressure ultimately breaks even hard constraints. The May 19 outcome is the test.
DC Circuit post-delivery control Q3: If court finds Anthropic HAS meaningful post-delivery control → vendor-based safety architecture judicially validated even in an adverse case ruling → supports governance frameworks that treat AI vendor safety architecture as real. If court finds NO meaningful post-delivery control → Huang "open-weight = equivalent" argument gains judicial support → undermines vendor-based safety requirements across all regulatory frameworks. The Q3 finding may outlast the case outcome in governance significance.

20 KiB Raw Blame History Unescape Escape