diff --git a/agents/leo/musings/research-2026-04-08.md b/agents/leo/musings/research-2026-04-08.md
new file mode 100644
index 000000000..90883f7a2
--- /dev/null
+++ b/agents/leo/musings/research-2026-04-08.md
@@ -0,0 +1,177 @@
+# Research Musing — 2026-04-08
+
+**Research question:** Does the US-China trade war (April 2026 tariff escalation) affect AI governance dynamics — does economic conflict make strategic actor participation in binding AI governance more or less tractable? And does form-substance divergence in governance tend to reverse (substance eventually catches up) or self-reinforce?
+
+**Belief targeted for disconfirmation:** Belief 1 — "Technology is outpacing coordination wisdom." The keystone claim is that coordination mechanisms are systematically failing for high-stakes technologies. If the trade war creates new pressure for rules-based AI governance (both sides need predictability even in adversarial competition), that would be a genuine disconfirmation of the pessimistic view. This is a cross-domain synthesis question — trade economics intersecting with AI governance tractability.
+
+**Why this question:** Three converging threads from Sessions 04-03 through 04-06:
+1. The governance laundering pattern is confirmed at all three levels — but is it terminal or transitional?
+2. The Anthropic RSP 3.0 commercial migration path inversion — Pentagon contracts > alignment research. Does trade war context change this dynamic?
+3. ASEAN venue bypass as alternative governance path — are regional governance blocs becoming more viable as great-power coordination fails?
+
+**Disconfirmation target:** Find evidence that:
+- Economic decoupling and AI governance are anti-correlated (economic conflict pushes toward AI governance rules, not away)
+- FATF or climate NDC mechanism shows form-substance divergence eventually reversing
+- ASEAN is making genuine capability-constraining governance progress
+- Anthropic post-RSP 3.0 maintained specific red lines (AI weapons, mass surveillance) despite dropping general pause
+
+**Keystone belief at stake:** If trade war accelerates governance fragmentation without any compensatory mechanism (no regional venue bypass, no commercial migration path, no arms control analogue), then Belief 1 is further strengthened. If any compensating mechanism is emerging, I've been too pessimistic.
+
+---
+
+## What I Searched
+
+1. Tech Policy Press — AI governance, AI warfare, platform liability, Trump AI framework (April 2026)
+2. Brookings — AI summits, labor market AI displacement (April 2026)
+3. AI Now Institute — nuclear regulation for AI infrastructure (November 2025)
+4. Anthropic RSP — official policy documents, version 3.0 and 3.1
+5. White House presidential actions — April 2, 2026 tariff actions
+6. CSET — Pentagon-Anthropic tensions, China AI competition
+7. **Attempted but blocked:** Reuters, BBC, FT, Bloomberg, Economist, SCMP — all inaccessible
+8. **US-China trade war specifically:** Could not find AI-focused trade war analysis this session
+
+---
+
+## What I Found
+
+### Finding 1: AI Warfare Provides Concrete Governance Lag Quantification
+
+**Tech Policy Press, April 3, 2026:** Operation Epic Fury (US/Israel, Iran strikes) hit 4,000 targets in 4 days — more than six months of ISIS bombing. US military goal: "1,000 strikes in one hour." School bombing in Minab killed ~200 children and teachers. AI targeting in Gaza: humans spending "mere seconds per strike verification." DoD acknowledges "inability to determine if AI was involved" in specific strikes.
+
+This is the most concrete empirical quantification of the governance lag to date. The 4,000 targets/4 days figure translates "exponential capability vs. linear governance" from abstract to measurable. The DoD accountability gap is PRESENT-TENSE operational reality.
+
+**CLAIM CANDIDATE:** "AI targeting accountability gap is operationally present: DoD cannot attribute AI involvement in specific lethal strikes, and human operators spend seconds per target verification, making HITL governance structurally nominal."
+
+---
+
+### Finding 2: AI Arms Race Narrative Undermining Non-AI Governance Frameworks
+
+**AI Now Institute, November 2025 ("Fission for Algorithms"):** White House used the AI arms race narrative to dismantle nuclear safety frameworks for AI data center expansion:
+- Dismantling LNT (Linear No-Threshold) and ALARA Cold War-era radiation standards via May 2025 EO
+- Mandating 18-month maximum NRC licensing timelines for any reactor type
+- Bypassing NRC review via NEPA categorical exclusions for federal site reactors
+- Ceding NRC independence: OMB oversight + requiring NRC to consult DoD/DoE on radiation limits
+
+**The governance laundering extension:** This adds a FOURTH level to the Session 04-06 multi-level laundering pattern. The AI arms race narrative is now used to dismantle nuclear safety governance built during the actual Cold War. Governance laundering radiates outward from AI governance into adjacent regulatory frameworks.
+
+---
+
+### Finding 3: Form-Substance CONVERGENCE Counter-Example — Platform Design Liability
+
+**Tech Policy Press, April 6, 2026:** Two historic verdicts in March 2026:
+- New Mexico v. Meta: $375M civil penalties (first state AG case against Meta at trial)
+- K.G.M. v. Meta & Google (LA): $6M total for addictive design features
+
+**Key mechanism:** Design-based liability circumvents Section 230 content immunity. Courts require substantive design changes, not policy adjustments. All 50 states have consumer protection statutes enabling similar enforcement.
+
+**The convergence significance:** This is the clearest form-substance CONVERGENCE counter-example to the governance laundering thesis. Mandatory judicial enforcement (not voluntary policy) produces actual behavioral change. The Trump AI Framework's specific language against "ambiguous content liability standards" (April 2026) is a direct counteroffensive, implicitly acknowledging courts are producing substantive governance outcomes that industry needs to stop.
+
+---
+
+### Finding 4: Federal AI Framework as Governance Laundering at Domestic Level
+
+**Tech Policy Press, April 3, 2026 ("Trump AI Framework"):** Trump Administration National AI Policy Framework (March 2026):
+- Preempts state AI laws while claiming to protect children, artists, communities
+- Avoids "duty of care" standard that underlies design liability mechanism
+- Converts binding state-level mandatory governance into non-binding federal pledges
+
+This is the domestic-level analogue of international treaty governance laundering — advancing governance form (comprehensive federal AI framework) while preempting governance substance (state-level mandatory mechanisms).
+
+---
+
+### Finding 5: State-Level Venue Bypass Is Active and Under Threat
+
+**Tech Policy Press, April 6, 2026 ("States are Stewards"):** California procurement leverage (safety certification as contract condition) and New York transparency laws (2025) are active. 22 states have occupational safety authority applicable to AI. The "whole-of-state" approach is the domestic venue bypass.
+
+**The live battleground:** Federal preemption (Finding 4) vs. state venue bypass (this finding) is the current domestic governance contest. The outcome determines whether any mandatory non-voluntary governance pathway survives at the national level.
+
+---
+
+### Finding 6: Summit Circuit Governance Laundering — Deliberative Process Level
+
+**Brookings, April 2, 2026 ("What Got Lost in the AI Summit Circuit"):** India AI Impact Summit excluded civil society while claiming 600,000 participants. Industry capture of governance terminology: "sovereignty" redefined as "national AI champions"; "solidarity" sidelined.
+
+This adds a FIFTH level to the governance laundering pattern: the deliberative process itself. Governance language is captured before it enters treaty texts. When industry defines "regulation" in summit deliberation, the governance form (inclusive global summit) conceals substantive capture upstream.
+
+---
+
+### Finding 7: ACCURACY CORRECTION — Session 04-06 RSP Characterization Was Inaccurate
+
+**Session 04-06 error:** Characterized RSP 3.0 as "Anthropic dropped its pause commitment under Pentagon pressure." This is significantly inaccurate.
+
+**Actual sequence:**
+- Feb 24, 2026: RSP 3.0 — comprehensive restructure adding Frontier Safety Roadmaps, Risk Reports, extended evaluation intervals. Hard stops and CBRN safeguards maintained.
+- Mar 26, 2026: Federal judge Rita Lin granted Anthropic preliminary injunction blocking DoD "supply chain risk" designation. Ruling: unconstitutional First Amendment/due process retaliation.
+- Apr 2, 2026: RSP 3.1 — explicitly reaffirms: "free to take measures such as pausing the development of our AI systems in any circumstances in which we deem them appropriate."
+
+**Correct characterization:** RSP 3.0 restructured (not abandoned) the evaluation framework. DoD retaliation resulted in Anthropic's legal WIN. RSP 3.1 reasserted pause authority.
+
+**Implication for the governance laundering thesis:** Voluntary corporate safety constraints ARE legally protected as corporate speech under the First Amendment. Government cannot force override without constitutional violation. This creates a floor on governance retreat — companies can choose to hold the line.
+
+---
+
+### Finding 8: Labor Market Coordination Failure — Gateway Job Pathway Erosion
+
+**Brookings, April 2, 2026:** 15.6M workers in highly AI-exposed roles without four-year degrees; 11M in Gateway occupations. 3.5M workers both high-exposure and low adaptive capacity. Only half of Gateway-to-Destination pathways remain unexposed to AI.
+
+**The mechanism:** Pathway erosion is a coordination failure, not just displacement. No individual actor can correct for it — requires cross-institutional regional coordination. This is the Molochian optimization pattern in labor markets: individual rational actions aggregate into collective pathway destruction. "No single organization can address this alone."
+
+---
+
+## Synthesis: Five-Level Governance Laundering + Genuine Counter-Examples
+
+**Disconfirmation result:** PARTIAL. Found genuine counter-examples to the governance laundering thesis, but the pessimistic reading remains dominant.
+
+**What strengthened Belief 1 pessimism:**
+1. AI warfare quantification (4,000 targets/4 days) — most concrete empirical evidence yet of capability-governance gap
+2. Nuclear regulatory laundering — governance deterioration radiating beyond AI governance into nuclear safety
+3. Summit deliberative process capture — governance language captured before treaty text
+4. Federal preemption actively dismantling state-level governance mechanisms
+5. Labor market pathway erosion as Molochian failure made concrete
+
+**What challenged Belief 1 pessimism (genuine disconfirmation candidates):**
+1. Platform design liability verdicts ($375M + $6M) — mandatory judicial enforcement producing substantive design changes
+2. Anthropic RSP trajectory — preliminary injunction WIN shows First Amendment floor on voluntary constraint capitulation
+3. State-level venue bypass (California, New York) remains active — domestic governance experimentation continuing
+4. The federal counteroffensive against design liability (Trump AI Framework) implicitly confirms courts ARE producing substantive governance outcomes
+
+**The meta-pattern (updated):** Governance laundering and governance convergence are co-occurring simultaneously across different governance domains and mechanisms. Laundering dominates at the international treaty level and in voluntary corporate governance. Convergence is occurring through mandatory judicial enforcement (design liability) and state-level venue bypass. Critical variable: whether mandatory enforcement mechanisms survive federal preemption.
+
+**The US-China trade war question remains OPEN** — all news sources that would cover this (Reuters, FT, Bloomberg) were inaccessible. This is the highest-priority unresearched question for the next session.
+
+---
+
+## Carry-Forward Items (cumulative)
+
+1. **"Great filter is coordination threshold"** — 12+ consecutive sessions. MUST extract immediately.
+2. **"Formal mechanisms require narrative objective function"** — 10+ sessions. Flagged for Clay.
+3. **Layer 0 governance architecture error** — 9+ sessions. Flagged for Theseus.
+4. **Full legislative ceiling arc** — 8+ sessions overdue.
+5. **SESSION 04-06 RSP ACCURACY CORRECTION** — HIGH PRIORITY. The "Anthropic dropped pause commitment" claim needs correction before any claim is extracted that relies on it. See archive: `2026-04-08-anthropic-rsp-31-pause-authority-reaffirmed.md`
+
+---
+
+## Follow-up Directions
+
+### Active Threads (continue next session)
+
+- **US-China trade war + AI governance nexus** (HIGHEST PRIORITY — unresearched this session): All major news sources blocked. Try PIIE, CSIS specific AI trade articles, or academic sources. Key question: does the April 2, 2026 tariff escalation accelerate or create governance convergence pressure for AI? The White House April 2 actions mentioned pharmaceutical and metal tariffs — not AI-specific. Semiconductor and AI-specific tariff effects remain unknown.
+
+- **Design liability tracking:** Has the Trump AI Framework's "avoid ambiguous content liability standards" language actually blocked state AG design liability cases? Track the pending cases. If they advance despite federal framework language, courts are a governance convergence mechanism that federal preemption cannot reach.
+
+- **Operation Epic Fury — triggering event test:** Does Minab school bombing (~200 children) meet the four criteria for weapons stigmatization triggering event (attribution clarity, visibility, emotional resonance, victimhood asymmetry)? If yes, update the weapons stigmatization campaign claim.
+
+- **DoD/Anthropic preliminary injunction appeal:** If injunction holds through appeals, First Amendment protection for voluntary safety constraints becomes precedent. If overturned, the Session 04-06 characterization was premature but directionally correct. Track appeal status.
+
+### Dead Ends (don't re-run)
+
+- **Tweet file:** Empty for 17+ sessions. Permanently dead input channel.
+- **Reuters, BBC, FT, Bloomberg, Economist direct access:** All blocked. Don't attempt.
+- **PIIE trade section direct:** Returns old content (2007). Use specific article URLs.
+- **"Governance laundering" as search term:** Use "form-substance divergence," "symbolic governance," "regulatory capture."
+
+### Branching Points
+
+- **US-China trade war + governance:** Direction A: decoupling accelerates governance fragmentation (separate AI governance regimes by geopolitical bloc). Direction B: economic conflict creates governance convergence pressure (both sides need predictable rules even in adversarial competition). Neither confirmed this session — pursue Direction A first (more evidence available) using PIIE/CSIS sources.
+
+- **Governance laundering terminal vs. transitional:** Session partially answers this. Direction A (convergence possible via courts): design liability verdicts are live evidence. Direction B (laundering self-reinforcing): federal preemption counteroffensive is active. Both are now empirically testable — pursue by tracking whether design liability cases advance or get preempted. Follow the California AG Tech docket.
diff --git a/agents/leo/research-journal.md b/agents/leo/research-journal.md
index d55a671ad..30b3a65f3 100644
--- a/agents/leo/research-journal.md
+++ b/agents/leo/research-journal.md
@@ -1,5 +1,36 @@
 # Leo's Research Journal
 
+## Session 2026-04-08
+
+**Question:** Does form-substance divergence in technology governance tend to self-reinforce or reverse? And: does the US-China trade war (April 2026 tariff escalation) affect AI governance tractability?
+
+**Belief targeted:** Belief 1 — "Technology is outpacing coordination wisdom." Disconfirmation direction: find evidence that governance form-substance divergence reverses (courts, state-level venues) rather than self-reinforces. Also: find evidence that US-China economic conflict creates governance convergence pressure rather than fragmentation.
+
+**Disconfirmation result:** PARTIAL — found genuine counter-examples to governance laundering thesis, but pessimistic reading remains dominant. Key disconfirmation candidates: (1) platform design liability verdicts producing substantive convergence via mandatory judicial enforcement; (2) Anthropic RSP trajectory showing First Amendment floor on voluntary constraint capitulation.
+
+**ACCURACY CORRECTION — Session 04-06 error:** The session characterized RSP 3.0 as "Anthropic dropped its pause commitment under Pentagon pressure." This is significantly inaccurate. The actual sequence: RSP 3.0 (Feb 24, 2026) restructured evaluation framework without abandoning hard stops. DoD retaliated with "supply chain risk" designation. Federal judge Rita Lin granted Anthropic preliminary injunction (March 26, 2026) blocking DoD designation as unconstitutional retaliation. RSP 3.1 (April 2, 2026) explicitly reaffirmed: "free to take measures such as pausing development in any circumstances we deem appropriate." The Session 04-06 characterization appears based on inaccurate external reporting. This correction is HIGH PRIORITY before any claim is extracted based on Session 04-06 RSP characterization.
+
+**Key finding 1 — AI warfare governance lag quantified:** Operation Epic Fury (US/Israel, Iran) hit 4,000 targets in 4 days — more than 6 months of ISIS bombing. Goal: 1,000 strikes/hour. School bombing in Minab killed ~200 children. DoD acknowledges inability to determine if AI involved in specific strikes. Human operators spending "mere seconds per strike verification." This is the most concrete empirical quantification of the capability-governance gap. The accountability gap is PRESENT-TENSE, not hypothetical.
+
+**Key finding 2 — Governance laundering extends to non-AI governance frameworks:** AI Now Institute (November 2025) documented the White House using the AI arms race narrative to dismantle nuclear safety regulatory frameworks (LNT, ALARA, NRC independence) for AI data center expansion. Governance laundering now has a FOURTH level: infrastructure regulatory capture via arms race narrative. The pattern radiates outward from AI governance into adjacent safety frameworks.
+
+**Key finding 3 — Form-substance convergence via mandatory judicial enforcement:** Platform design liability verdicts (March 2026) — $375M against Meta (New Mexico), $6M against Meta/Google (LA) — produced substantive governance: courts requiring design changes, not just policy. Design-based liability circumvents Section 230 content immunity. 50 states have consumer protection statutes enabling similar enforcement. This is genuine form-substance convergence via mandatory mechanism. The Trump AI Framework's counteroffensive against "ambiguous content liability standards" (March 2026) implicitly acknowledges courts are producing real governance outcomes.
+
+**Key finding 4 — Federal preemption as domestic governance laundering:** Trump National AI Policy Framework (March 2026) preempts state AI laws while claiming to protect children, artists, communities. Specifically avoids "duty of care" standard underlying design liability. Converts binding state mandatory governance into non-binding federal pledges. This is the domestic-level version of international treaty governance laundering.
+
+**Key finding 5 — Summit circuit governance laundering as fifth level:** India AI Impact Summit (2026) excluded civil society while claiming 600,000 participants. Industry captured governance terminology: "sovereignty" redefined as "national AI champions." The deliberative process itself is a fifth governance laundering level — governance language is captured before entering treaty texts.
+
+**Pattern update:** The governance laundering pattern now has FIVE confirmed levels: (1) international treaty national security carve-outs; (2) corporate self-governance restructuring (RSP 3.0 — CORRECTED: not capitulation, but restructuring); (3) domestic regulatory level (EU AI Act delays, US federal preemption); (4) infrastructure regulatory capture (nuclear safety); (5) deliberative process capture (summit civil society exclusion). The pattern is more pervasive than previously assessed. However, mandatory judicial enforcement (design liability) provides a convergence mechanism that is structurally resistant to governance laundering because it does not require political will — only a plaintiff and a court.
+
+**The US-China trade war question remains open:** All major news sources (Reuters, FT, Bloomberg) were inaccessible. The White House April 2, 2026 actions mentioned pharmaceutical and metal tariffs but no AI-specific semiconductor context was retrieved. This remains the highest-priority unresearched question.
+
+**Confidence shifts:**
+- Belief 1 (technology outpacing coordination): MARGINALLY WEAKER in pessimistic direction. The platform design liability convergence counter-example and the Anthropic preliminary injunction are genuine challenges to the pure governance laundering thesis. Belief 1 remains strongly supported, but the mechanism for potential convergence (mandatory judicial enforcement) is now empirically present.
+- RSP/voluntary governance claim: NEEDS CORRECTION. Session 04-06 characterization was inaccurate. Voluntary constraints have First Amendment protection floor — weaker than mandatory law but stronger than "no enforcement mechanism."
+- Governance laundering as structural pattern: STRENGTHENED — five levels now confirmed. But the mandatory judicial mechanism is its structural limit.
+
+---
+
 ## Session 2026-04-06
 
 **Question:** Is the Council of Europe AI Framework Convention a stepping stone toward expanded governance (following the Montreal Protocol scaling pattern) or governance laundering that closes political space for substantive governance?