agents/theseus/musings/research-2026-03-18.md

@@ -117,23 +117,99 @@ The net effect is time-dependent, and economic forces optimize for the SHORT ter
 
 Total: 8 sources (7 high, 1 medium)
 
+---
+
+## Session 2: Correction Mechanisms (2026-03-18, continuation)
+
+**Research question:** What correction mechanisms could address the systematic automation overshoot identified in Session 1?
+
+**Disconfirmation target:** If effective governance or market mechanisms exist that correct for overshoot, the "not being treated as such" component of keystone belief B1 weakens.
+
+### Finding 6: Four correction mechanism categories exist — all have a shared structural limitation
+
+**Market-based — AI liability insurance (AIUC/Munich Re):**
+AIUC launched the world's first AI agent certification (AIUC-1) in July 2025, covering six pillars: security, safety, reliability, data/privacy, accountability, societal risks. Insurance market projected at ~$4.7B by 2032. Mechanism: insurers profit from accurately pricing risk → financial incentive to measure outcomes accurately → coverage contingent on safety standards → pre-market safety pressure. Historical precedent is strong: fire insurance → building codes (Franklin); seatbelt adoption driven partially by insurance premium incentives. Munich Re: "insurance has played a major role in [safety improvements], and I believe insurance can play the same role for AI."
+
+**Regulatory — EU AI Act Article 14 (enforcement August 2026):**
+Mandatory human oversight with competency and training requirements for high-risk AI systems. Key provisions: (a) natural persons with "necessary competence, training and authority" must be assigned to oversight; (b) for highest-risk applications, no action taken unless SEPARATELY VERIFIED AND CONFIRMED by at least two natural persons. Training programs must cover AI capabilities AND limitations, risk awareness, and intervention procedures. The two-person verification rule is structurally notable — it's a mandatory human-in-the-loop requirement that prevents single-point override.
+
+**Organizational — Reliance drills and analog practice (Hosanagar/Wharton):**
+Proposed by analogy to aviation: FAA now mandates manual flying practice after Air France 447 (autopilot deskilling → crash). AI equivalent: "off-AI days" and failure scenario stress tests. Individual-level: require human first drafts before AI engagement; build deliberate review checkpoints. The FAA aviation case is significant: government mandated the intervention after a catastrophic failure. Deskilling correction required regulatory forcing, not voluntary adoption.
+
+**Cryptoeconomic — Agentbound Tokens (Chaffer/McGill, working paper):**
+ABTs apply Taleb's skin-in-the-game to AI agents: staking collateral to access high-risk tasks, automatic slashing for misconduct, reputation decay. Design principle: "accountability scales with autonomy." Decentralized validator DAOs (human + AI hybrid). Per-agent caps prevent monopolization. Most theoretically elegant mechanism found — addresses the accountability gap directly without government coordination. Currently: working paper, no deployment.
+
+### Finding 7: All four mechanisms share a measurement dependency — the perception gap corrupts them at the source
+
+This is the session's key insight. Every correction mechanism requires accurate outcome measurement to function:
+- Insurance requires reliable claims data (can't price risk if incidents aren't reported or recognized)
+- EU AI Act compliance requires evidence of actual oversight capability (not just stated)
+- Reliance drills require knowing when capability has eroded (can't schedule them if you can't detect the erosion)
+- ABTs require detecting misconduct (slashing only works if violations are observable)
+
+But the METR RCT (Session 1, Mechanism 1) showed a 39-point gap between perceived and actual AI benefit. This is a SELF-ASSESSMENT BIAS that corrupts the measurement signals all correction mechanisms depend on. This creates a second-order market failure: mechanisms designed to correct the first failure (overshoot) themselves fail because the information that would trigger them is unavailable or biased.
+
+Automation bias literature (2025 systematic review, 35 studies) provides the cognitive mechanism: nonlinear relationship between AI knowledge and reliance. The "Dunning-Kruger zone" — small exposure → overconfidence → overreliance — is where most enterprise adopters sit. Conditions that DRIVE AI adoption (high workload, time pressure) are the SAME conditions that MAXIMIZE automation bias. Self-reinforcing feedback loop at the cognitive level.
+
+### Finding 8: AI's economic value is being systematically misidentified — misallocation compounds overshoot
+
+HBR/Choudary (Feb 2026): AI's actual economic payoff is in reducing "translation costs" — friction in coordinating disparate teams, tools, and data — not in automating individual tasks. AI enables coordination WITHOUT requiring consensus on standards or platforms (historically the barrier). Examples: Tractable disrupted CCC by interpreting smartphone photos without standardization; Trunk Tools integrates BIM, spreadsheets, photos without requiring all teams to switch platforms.
+
+If correct, this means most AI deployment (automation-focused) is optimizing for the LOWER-VALUE application. Organizations are overshooting automation AND underinvesting in coordination. This is a value misallocation that compounds the overshoot problem: not only are firms using more AI than is optimal for automation, they're using it for the wrong thing.
+
+This connects directly to our KB coordination thesis: if AI's value is in coordination reduction, then AI safety framing should also be coordination-first. The argument is recursive.
+
+### Finding 9: Government as coordination-BREAKER confirmed with specific episode
+
+HKS/Carr-Ryan Center (2026): The DoD threatened to blacklist Anthropic unless it removed safeguards against mass surveillance and autonomous weapons. Anthropic refused publicly; Pentagon retaliated. Critical implication: "critical protections depend entirely on individual corporate decisions rather than binding international frameworks." CFR confirms: "large-scale binding international agreements on AI governance are unlikely in 2026" (Horowitz). Governance happening through bilateral government-company negotiations "without transparency, without public accountability, and without remedy mechanisms."
+
+This is not a peripheral data point. This is the government functioning as a coordination-BREAKER — actively penalizing safety constraints — rather than a correction mechanism. Extends and updates the existing KB claim about [[government designation of safety-conscious AI labs as supply chain risks]].
+
+### Disconfirmation result (B1 keystone belief)
+
+**Verdict:** Partial disconfirmation. More correction mechanisms exist than I was crediting (AIUC-1 certification is real, EU AI Act Art 14 is real, ABT framework is published). WEAKENS the "not being treated as such" component in degree but not in direction.
+
+**Offset factors:** 63% of organizations lack AI governance policies (IBM/Strategy International); binding international agreements "unlikely in 2026"; government is functioning as coordination-BREAKER (DoD/Anthropic); EU AI Act covers only "high-risk" defined systems, not general enterprise deployment; all mechanisms share measurement dependency that the perception gap corrupts. The gap between severity and response remains structurally large.
+
+**Net confidence shift on B1:** Belief holds. "Not being treated as such" is still accurate at the level of magnitude of response vs. magnitude of risk. The mechanisms being built are real but mismatched in scale.
+
+### The Missing Mechanism
+
+No existing correction mechanism addresses the perception gap directly. All four categories are SECOND-ORDER mechanisms (they require information the first-order failure corrupts). The gap: mandatory, standardized, THIRD-PARTY performance measurement before and after AI deployment — not self-reported, not self-assessed, independent of the deploying organization. This would create the information basis that all other mechanisms depend on.
+
+Analogy: drug approval requires third-party clinical trials, not manufacturer self-assessment. Aviation safety requires flight data recorder analysis, not pilot self-report. AI adoption currently has no equivalent. This is the gap.
+
+## Sources Archived This Session (Session 2)
+
+1. **Hosanagar (Substack) — AI Deskilling Prevention** (HIGH) — reliance drills, analog practice, FAA analogy
+2. **NBC News/AIUC — AI Insurance as Safety Mechanism** (HIGH) — AIUC-1 certification, market-based correction, Munich Re
+3. **Chaffer/McGill — Agentbound Tokens** (MEDIUM) — cryptoeconomic accountability, skin-in-the-game
+4. **Choudary/HBR — AI's Big Payoff Is Coordination** (HIGH) — translation costs, coordination vs. automation value
+5. **HKS Carr-Ryan — Governance by Procurement** (HIGH) — bilateral negotiation failure, DoD/Anthropic episode
+6. **Strategy International — Investment Outruns Oversight** (MEDIUM) — $405B/$650B investment data, 63% governance deficit
+
+Total Session 2: 6 sources (4 high, 2 medium)
+Total across both sessions: 14 sources
+
 ## Follow-up Directions
 
 ### NEXT: (continue next session)
-- **Formal characterization of overshoot dynamics**: The four mechanisms need a unifying formal model. Is this a market failure (externalities), a principal-agent problem (perception gap), a commons tragedy (collective intelligence as commons), or something new? The framework matters for what interventions would work. Search for: economic models of technology over-adoption, Jevons paradox applied to AI, rebound effects in automation.
-- **Correction mechanisms that could work**: If self-correction fails (perception gap) and market forces overshoot (competitive pressure), what coordination mechanisms could maintain optimal integration? Prediction markets on team performance? Mandatory human-AI joint testing (JAT framework)? Regulatory minimum human competency requirements? This connects to Rio's mechanism design expertise.
-- **Temporal dynamics of the inverted-U peak**: Finding 3 shows diversity increasing over time in hybrids. Finding 4 shows homogenization eroding human diversity. These are opposing forces. Does the peak move UP (as hybrid networks learn) or DOWN (as homogenization erodes inputs)? This needs longitudinal data.
+- **Third-party performance measurement infrastructure**: The missing correction mechanism. What would mandatory independent AI performance assessment look like? Who would run it? Aviation (FAA flight data), pharma (FDA trials), finance (SEC audits) all have equivalents. Is there a regulatory proposal for AI equivalent? Search: "AI performance audit" "third-party AI assessment" "mandatory AI evaluation framework" 2026.
+- **Formal characterization of overshoot dynamics**: The four mechanisms still need unifying formal model. Market failure taxonomy: externalities (competitive pressure), information failure (perception gap), commons tragedy (collective intelligence as commons), bounded rationality (verification tax). Are these all the same underlying mechanism or distinct? Jevons paradox applied to AI: does AI use expand to fill saved time?
+- **Temporal dynamics of inverted-U peak**: Finding 3 (diversity increases over time in hybrids) vs. Finding 4 (homogenization erodes human diversity). These are opposing forces. Longitudinal data needed.
 
 ### COMPLETED: (threads finished)
-- **"Does economic force push past optimal?"** — YES, through four independent mechanisms. The open question from _map.md is answered: the net effect is time-dependent, and economic forces optimize for the wrong time horizon.
-- **Session 5 (2026-03-12) incomplete musing** — This session completes that research question with substantial evidence.
+- **Correction mechanisms question** — answered: four categories exist (market, regulatory, organizational, cryptoeconomic), all share measurement dependency. Missing mechanism identified: third-party performance measurement.
+- **Keystone belief disconfirmation search** — completed: mechanisms more developed than credited, but gap between severity and response remains structurally large. B1 holds.
 
 ### DEAD ENDS: (don't re-run)
-- ScienceDirect, Cell Press, Springer, CACM, WEF, CNBC all blocked by paywalls/403s via WebFetch
-- "Verification tax" as a search term returns tax preparation AI, not the concept — use "AI verification overhead" or "hallucination mitigation cost" instead
+- WEF, Springer (Springer gave 303 redirect), Nature (Science Reports), PMC (reCAPTCHA) all blocked
+- ScienceDirect, Cell Press, CACM still blocked (from Session 1)
+- "Prediction markets AI governance" search returns enterprise AI predictions, not market mechanisms for governance — use "mechanism design AI accountability" or "cryptoeconomic AI safety" instead
 
 ### ROUTE: (for other agents)
-- **Seven feedback loops (L1-L7)** → **Rio**: The competitive adoption cycle is the alignment tax applied to economic decisions. The demand destruction loop (adoption → displacement → reduced consumer income → demand destruction) is a market failure that prediction markets or mechanism design might address.
-- **Seven feedback loops (L7)** → **Leo**: The time-compression meta-crisis (exponential technology vs linear governance) directly confirms Leo's coordination thesis and deserves synthesis treatment.
-- **AI homogenization of expression** → **Clay**: If AI is standardizing how people write and think, this directly threatens narrative diversity — Clay's territory. The social pressure mechanism (conform to AI-standard communication) is a cultural dynamics claim.
-- **Deskilling evidence** → **Vida**: Endoscopist deskilling (28.4% → 22.4% detection rate) is medical evidence Vida should evaluate. The self-reinforcing loop applies to clinical AI adoption decisions.
+- **AI insurance mechanism** → **Rio**: AIUC-1 certification + Munich Re involvement = market-based safety mechanism. Is this analogous to a prediction market? The certification requirement creates a skin-in-the-game structure Rio should evaluate.
+- **Agentbound Tokens (ABTs)** → **Rio**: Cryptoeconomic staking, slashing, validator DAOs. This is mechanism design for AI agents — Rio's expertise. The "accountability scales with autonomy" principle may generalize beyond AI to governance mechanisms broadly.
+- **HBR/Choudary translation costs** → **Leo**: If AI's value is in coordination reduction (not automation), this has civilizational implications for how we should frame AI's role in grand strategy. Leo should synthesize.
+- **DoD/Anthropic confrontation** → **Leo**: Government-as-coordination-BREAKER is a grand strategy claim — the state monopoly on force interacting with AI safety. Leo should evaluate whether this changes the [[nation-states will inevitably assert control]] claim.
+- **Bilateral governance failure** → **Rio**: Bilateral government-company AI negotiations = no transparency, no remedy mechanisms. Is there a market mechanism that could substitute for the missing multilateral governance? Prediction markets on AI safety outcomes?

agents/theseus/research-journal.md

@@ -173,3 +173,35 @@ NEW PATTERN:
 **Sources archived:** 8 sources (7 high, 1 medium). Key: Vaccaro et al. Nature HB meta-analysis, METR developer RCT, Sourati et al. Trends in Cognitive Sciences, EU AI Alliance seven feedback loops, collective creativity dynamics (arxiv), Forrester verification tax data, AI Frontiers high-stakes degradation, MIT Sloan J-curve.
 
 **Cross-session pattern (6 sessions):** Session 1 → theoretical grounding (active inference). Session 2 → empirical landscape (alignment gap bifurcating). Session 3 → constructive mechanisms (bridging, MaxMin, pluralism). Session 4 → mechanism engineering + complication (homogenization threatens diversity). Session 5 → [incomplete]. Session 6 → automation overshoot confirmed with four mechanisms. The progression: WHAT → WHERE → HOW → BUT ALSO → [gap] → WHY IT OVERSHOOTS. Next session should address: correction mechanisms — what coordination infrastructure prevents overshoot? This connects to Rio's mechanism design (prediction markets on team performance?) and our collective architecture (does domain specialization naturally prevent homogenization?).
+
+## Session 2026-03-18b (Correction Mechanisms)
+
+**Question:** What correction mechanisms could address systematic automation overshoot — and do their existence weaken the keystone belief that alignment is "not being treated as such"?
+
+**Belief targeted:** B1 (keystone) — "AI alignment is the greatest outstanding problem for humanity and not being treated as such." Specifically the disconfirmation target: do effective governance mechanisms keep pace with capability advances?
+
+**Disconfirmation result:** Partial disconfirmation. More correction mechanisms exist than previously credited: AIUC-1 AI agent certification (July 2025), EU AI Act Article 14 mandatory human competency requirements (enforcement August 2026), Agentbound Tokens cryptoeconomic accountability (working paper), organizational reliance drills (Hosanagar/Wharton). Each is real. BUT: all four share a measurement dependency the perception gap corrupts. 63% of organizations lack AI governance policies; binding international agreements "unlikely in 2026" (CFR/Horowitz); DoD threatened to blacklist Anthropic for maintaining safety safeguards. Net: mechanisms are more developed than credited, but the gap between severity and response remains structurally large.
+
+**Key finding:** All correction mechanisms share a second-order market failure: they require accurate outcome measurement to function, but the perception gap (METR RCT: 39-point gap) corrupts that information at the source. Insurance needs reliable claims data; regulation needs compliance evidence; organizational drills need to detect capability erosion; cryptoeconomic slashing needs to detect misconduct. The missing mechanism is third-party independent performance measurement — the equivalent of FDA clinical trials or aviation flight data recorders for AI deployment.
+
+**Pattern update:**
+
+STRENGTHENED:
+- B1 (alignment not being treated as such) — holds. Mechanisms exist but are mismatched in scale to the severity of the problem. The DoD/Anthropic confrontation is a concrete case of government functioning as coordination-BREAKER.
+- B2 (alignment is a coordination problem) — automation overshoot correction is also a coordination failure. The four mechanisms require coordination across firms/regulators to function; firms acting individually cannot correct for competitive pressure.
+- "Government as coordination-breaker" — updated with DoD/Anthropic episode. This is a stronger confirmation of the [[government designation of safety-conscious AI labs as supply chain risks]] claim.
+
+COMPLICATED:
+- The measurement dependency insight complicates all constructive alternatives. Even if we build collective intelligence infrastructure (B5), it needs accurate performance signals to self-correct. The perception gap at the organizational level is a precursor problem that the constructive case hasn't addressed.
+
+NEW PATTERN:
+- **Misallocation compounds overshoot.** HBR/Choudary (Feb 2026): AI's actual payoff is in reducing translation costs (coordination), not automating tasks. Most deployment is automation-focused. So firms are both OVER-ADOPTING AI for lower-value applications AND UNDER-ADOPTING for higher-value coordination. Two simultaneous misallocations, working in opposite directions on a single deployment trajectory.
+- **AI perception gap has a cognitive mechanism.** 2025 systematic review of automation bias (35 studies): Dunning-Kruger pattern — small AI exposure → overconfidence → overreliance. Conditions that drive adoption (time pressure, high workload) are the same conditions that maximize automation bias. Second self-reinforcing loop at the cognitive level.
+
+**Confidence shift:**
+- "Correction mechanisms are largely absent" → REVISED: mechanisms exist but all have measurement dependency. Better framing: "four correction mechanism categories exist but share a structural second-order failure."
+- "AI's economic value is in coordination not automation" → NEW, likely, based on HBR/Choudary analysis and consistent with coordination protocol > model scaling evidence
+- "Government as coordination-breaker is systematic" → UPDATED: DoD/Anthropic episode adds specific 2026 evidence
+- Keystone belief B1: unchanged in direction, weakened slightly in magnitude of the "not being treated as such" claim
+
+**Cross-session pattern (7 sessions):** Active inference → alignment gap → constructive mechanisms → mechanism engineering → [gap] → overshoot mechanisms → correction mechanism failures. The progression through this entire arc: WHAT our architecture should be → WHERE the field is → HOW specific mechanisms work → BUT ALSO mechanisms fail → WHY they overshoot → HOW correction fails too. The emerging thesis: the problem is not that solutions don't exist — it's that the INFORMATION INFRASTRUCTURE to deploy solutions is missing. Third-party performance measurement is the gap. Next: what would that infrastructure look like, and who is building it?

inbox/queue/2025-01-00-chaffer-agentbound-tokens-ai-accountability.md

@@ -0,0 +1,65 @@
+---
+type: source
+title: "Can We Govern the Agent-to-Agent Economy? Agentbound Tokens as Accountability Infrastructure"
+author: "Tomer Jordi Chaffer"
+url: https://arxiv.org/html/2501.16606v2
+date: 2025-01-01
+domain: ai-alignment
+secondary_domains: [internet-finance]
+format: article
+status: unprocessed
+priority: medium
+tags: [agentbound-tokens, accountability, skin-in-the-game, cryptoeconomics, mechanism-design, AI-agents, governance]
+flagged_for_rio: ["Cryptoeconomic mechanism design for AI agent accountability — tiered staking, slashing, DAO governance. Rio should evaluate whether the staking mechanism has prediction market properties for surfacing AI reliability signals"]
+---
+
+## Content
+
+**Agentbound Tokens (ABTs):** Cryptographic tokens serving as "tamper-proof digital birth certificates" for autonomous AI agents. Immutable identity markers that evolve dynamically based on agent performance and ethical compliance.
+
+**Core mechanism (skin-in-the-game):**
+- Agents stake ABTs as collateral to access high-risk tasks
+- Misconduct triggers automatic token slashing (proportional penalty)
+- Example: trading AI locks "market-compliant" ABT to access stock exchange data; manipulative trading → automatic token slash
+- Temporary blacklisting for repeat offenses
+- Delegated authority: agents can lease credentials while retaining liability
+
+**Accountability infrastructure:**
+- Dynamic credentialing reflecting ongoing compliance
+- Automated penalty systems (proportional to violation severity)
+- Decentralized validator DAOs (human + AI hybrid oversight)
+- Utility-weighted governance: governance power derives from verifiable utility to ecosystem (task success rates, energy efficiency), not just token quantity
+- Per-agent caps prevent monopolization
+- Reputation decay discourages hoarding
+
+**Key design principle:** "Accountability scales with autonomy" — higher autonomy requires higher stake
+
+**Author:** Tomer Jordi Chaffer (McGill University), with contributions from Goldston, Muttoni, Zhao, Shaw Walters. Working paper.
+
+## Agent Notes
+
+**Why this matters:** ABTs operationalize Taleb's skin-in-the-game principle for AI agents with specificity. The staking-and-slashing mechanism creates consequences that are: (a) automatic (no human discretion needed), (b) proportional (stakes scale with autonomy), (c) decentralized (validator DAOs, not single regulator). This is theoretically the most elegant correction mechanism found because it addresses the accountability gap directly without requiring government coordination.
+
+**What surprised me:** The "accountability scales with autonomy" principle is a clean solution to a genuine design problem — most governance proposals treat accountability as binary. Also: the DAO governance model includes both human and AI validators, which is closer to our collective superintelligence architecture than any governance proposal I've seen.
+
+**What I expected but didn't find:** Empirical validation — this is a working paper with no deployed system. Also: the mechanism assumes reliable outcome measurement (know when misconduct occurred), which runs into the perception gap problem again. The slashing mechanism only works if misconduct is detectable.
+
+**KB connections:**
+- [[multipolar failure from competing aligned AI systems may pose greater existential risk than any single misaligned superintelligence]] — ABTs are one mechanism for governing multi-agent interaction without requiring consensus
+- [[no research group is building alignment through collective intelligence infrastructure]] — this paper is evidence of early infrastructure-building, though at working-paper stage
+- [[coding agents cannot take accountability for mistakes]] — ABTs are a direct proposed solution to this claim
+
+**Extraction hints:**
+- Claim candidate: "cryptoeconomic staking mechanisms can create accountability for AI agents because automatic token slashing makes misconduct costly without requiring human discretionary oversight"
+- Critical limitation: only corrects DETECTABLE misconduct. Does not address the perception gap or coordination failures that operate at organizational level rather than agent level.
+- The "accountability scales with autonomy" principle may be extractable as a design principle, independent of the ABT implementation.
+
+**Context:** Working paper from McGill researcher — not peer reviewed. Cryptoeconomic framing will be familiar to Rio. Mechanism is theoretically grounded but empirically untested.
+
+## Curator Notes
+
+PRIMARY CONNECTION: [[coding agents cannot take accountability for mistakes which means humans must retain decision authority over security and critical systems regardless of agent capability]]
+
+WHY ARCHIVED: First governance mechanism specifically designed for AI agent accountability using cryptoeconomic principles. Also relevant to Rio's mechanism design territory.
+
+EXTRACTION HINT: Focus on the accountability-scales-with-autonomy principle and the staking model structure. Note the key limitation: measurement dependency. Do not over-claim — this is a working paper with no deployment evidence.

inbox/queue/2026-02-00-choudary-hbr-ai-coordination-not-automation.md

@@ -0,0 +1,58 @@
+---
+type: source
+title: "AI's Big Payoff Is Coordination, Not Automation"
+author: "Sangeet Paul Choudary (@sanguit)"
+url: https://hbr.org/2026/02/ais-big-payoff-is-coordination-not-automation
+date: 2026-02-01
+domain: ai-alignment
+secondary_domains: []
+format: article
+status: unprocessed
+priority: high
+tags: [coordination, automation, translation-costs, AI-value, misallocation, platform-strategy, economic-payoff]
+---
+
+## Content
+
+**Main argument:** AI's most significant economic value comes from reducing "translation costs" — friction in coordinating disparate teams, tools, and data — rather than automating individual tasks. AI enables coordination without requiring consensus on standards or platforms.
+
+**Key concept — Translation costs:** The friction involved in coordinating disparate teams, tools, systems. Historically required standardization (everyone use the same platform). AI eliminates the standardization requirement by doing the translation dynamically.
+
+**Evidence:**
+- **Construction (Trunk Tools):** Integrates BIM software, spreadsheets, photos, emails, PDFs into unified project view. Teams maintain specialized tools. Coordination cost drops without standardization.
+- **Auto insurance (Tractable):** Disrupted market leader CCC Intelligent Solutions by training AI to interpret smartphone photos of vehicle damage — sidestepping standardization requirements. Processed ~$7B in claims by 2023.
+
+**Author's three strategies for incumbents:**
+1. Become the translation layer (example: project44 in logistics — ecosystem-wide coordination)
+2. Double down on accountability (Maersk's integrated logistics model — responsible for outcomes despite fragmentation)
+3. Fragment and tax (FedEx — maintains privileged internal unified view, rations external access)
+
+**Author:** Sangeet Paul Choudary — C-level AI and platform strategy advisor, UC Berkeley senior fellow, Thinkers50 Strategy Award 2025.
+
+## Agent Notes
+
+**Why this matters:** This is the most important reframe I've encountered for the automation overshoot problem. If AI's ACTUAL value is in coordination reduction (not automation), then organizations that are automating tasks (the dominant deployment pattern) are SYSTEMATICALLY MISALLOCATING. They're pursuing the wrong value. This is a new mechanism for misallocation that's distinct from the four overshoot mechanisms identified last session — it's not that firms overshoot the optimal automation level, it's that they're optimizing for the wrong thing entirely.
+
+**What surprised me:** The argument that AI eliminates the standardization requirement for coordination is genuinely novel to me. This matches the mathematical argument in our KB — distributed architectures don't require consensus (like monolithic alignment trying to aggregate all preferences). If AI can coordinate without consensus, this is a practical instantiation of what our collective architecture thesis requires theoretically.
+
+**What I expected but didn't find:** Evidence that the coordination payoff is LARGER than automation in magnitude. The article makes the qualitative argument but doesn't provide comparative ROI data. Also missing: whether coordination applications of AI are being deployed at scale yet, or whether this remains largely untapped.
+
+**KB connections:**
+- [[coordination protocol design produces larger capability gains than model scaling]] — directly confirmed: coordination > automation as the value driver
+- [[AI alignment is a coordination problem not a technical problem]] — if AI's VALUE is in coordination, then AI SAFETY must also be framed as coordination (recursive alignment of the argument)
+- [[collective intelligence is a measurable property of group interaction structure not aggregated individual ability]] — AI reducing translation costs IS improving group interaction structure
+
+**Extraction hints:**
+- High-priority claim candidate: "AI's primary economic value is in reducing translation costs between specialized teams and tools rather than automating individual tasks, which means most AI deployment is systematically misallocated toward lower-value automation applications"
+- The "coordination without consensus" principle deserves extraction — it operationalizes the distributed architecture thesis at the firm level
+- The three incumbent strategies are less extractable (prescriptive rather than empirical)
+
+**Context:** HBR February 2026 publication by credible platform strategy thinker. Highly visible to business audience. This is the kind of mainstream articulation that could shift how organizations think about AI deployment.
+
+## Curator Notes
+
+PRIMARY CONNECTION: [[coordination protocol design produces larger capability gains than model scaling because the same AI model performed 6x better with structured exploration than with human coaching on the same problem]]
+
+WHY ARCHIVED: Provides the economic theory for WHY automation-focused AI deployment is suboptimal — the real value is in coordination. This reframes the overshoot problem as misallocation not just excess.
+
+EXTRACTION HINT: Extract the "translation costs" concept and the coordination-vs-automation value claim. Scope carefully: Choudary's argument is about where economic value is largest, not about alignment implications — Theseus should make the alignment connection explicit in extraction.

inbox/queue/2026-02-00-hosanagar-ai-deskilling-prevention-interventions.md

@@ -0,0 +1,61 @@
+---
+type: source
+title: "AI Is Deskilling You. Here's How to Prevent It"
+author: "Kartik Hosanagar (@kartikh)"
+url: https://hosanagar.substack.com/p/ai-is-deskilling-you-heres-how-to
+date: 2026-02-01
+domain: ai-alignment
+secondary_domains: [health]
+format: article
+status: unprocessed
+priority: high
+tags: [deskilling, human-competency, reliance-drills, analog-practice, automation-overshoot, organizational-intervention]
+---
+
+## Content
+
+Author (Wharton professor): AI deskilling is real and requires deliberate organizational intervention. Three case studies:
+
+**Aviation:** 2009 Air France 447 crash — pilots lost manual flying skills through automation dependency. FAA now requires mandatory manual practice sessions.
+
+**Medicine:** Endoscopists using AI for polyp detection became worse at finding polyps when AI was turned off. Adenoma detection dropped from 28% to 22% without AI (same data as Lancet Gastroenterology cited in previous sessions).
+
+**Education:** Students with unrestricted GPT-4 access initially performed better at math, but underperformed compared to peers who never used AI once access was removed.
+
+**Proposed interventions:**
+
+Individual level:
+- Practice "mindful" AI use — distinguish between skills deliberately outsourced vs. skills being eroded
+- Require human first rounds (sketches, assumptions, hypotheses) before AI assistance
+- Build deliberate review points to re-engage judgment
+
+Organizational level:
+- **Reliance Drills**: Routine stress tests simulating AI failure or unavailability — expose knowledge erosion before crises. E.g., failure scenarios where teams reach decisions without AI, or "off-AI days"
+- **Analog Practice**: Required independent thinking and creation to maintain resilience; analogous to pilots' mandatory manual flying requirements
+
+## Agent Notes
+
+**Why this matters:** Provides specific, actionable organizational interventions for preventing the deskilling drift that was identified as Mechanism 3 of automation overshoot. The reliance drills concept is directly analogous to how aviation solved its equivalent problem — and aviation solved it through regulatory mandate (FAA). This suggests the deskilling correction mechanism requires regulatory forcing, not voluntary adoption.
+
+**What surprised me:** The three-domain evidence convergence (aviation → medicine → education) across independent fields all showing the same deskilling pattern makes this much stronger than any single-domain claim. The FAA mandate for manual practice is the closest analogue I've found to what a regulatory correction mechanism for AI deskilling would look like.
+
+**What I expected but didn't find:** Specific evidence that reliance drills or analog practice work in AI contexts — these are proposed by analogy, not yet tested. The aviation fix took decades after the problem was identified. The organizational interventions remain voluntary and self-selected.
+
+**KB connections:**
+- [[AI capability and reliability are independent dimensions]] — deskilling is the human-side version of this problem
+- [[human-in-the-loop clinical AI degrades to worse-than-AI-alone]] — same mechanism, different direction
+- [[economic forces push humans out of every cognitive loop]] — the economic force the author is trying to correct against
+
+**Extraction hints:**
+- Claim candidate: "reliance drills and analog practice are the minimum viable organizational intervention for preventing AI deskilling because they create the regular human-independent practice that historically has prevented capability erosion in other high-stakes domains"
+- Could also extract: "FAA mandatory manual flying requirements are the regulatory template for AI deskilling prevention in high-stakes domains"
+
+**Context:** Hosanagar is a credible Wharton academic with AI expertise. The Substack format means this is less formally reviewed than his academic work, but the argument is empirically grounded.
+
+## Curator Notes
+
+PRIMARY CONNECTION: [[economic forces push humans out of every cognitive loop where output quality is independently verifiable]] (the force these interventions push back against)
+
+WHY ARCHIVED: First source with specific, concrete organizational interventions against deskilling drift — the third overshoot mechanism. Also provides the FAA regulatory template analogy.
+
+EXTRACTION HINT: Extractor should focus on (a) the reliance drills concept as a claim about minimum viable organizational intervention, and (b) FAA mandatory practice as regulatory template. Do not extract the case studies — those are already in KB from other sources.

inbox/queue/2026-03-11-strategyinternational-ai-investment-outruns-oversight.md

@@ -0,0 +1,64 @@
+---
+type: source
+title: "AI at Scale: When Investment Outruns Oversight"
+author: "Strategy International Think Tank"
+url: https://strategyinternational.org/2026/03/11/publication252/
+date: 2026-03-11
+domain: ai-alignment
+secondary_domains: []
+format: article
+status: unprocessed
+priority: medium
+tags: [investment, oversight, governance-deficit, deployment-pressure, AI-scale, accountability]
+---
+
+## Content
+
+**Core argument:** Massive capital investments in AI infrastructure are creating pressure to deploy systems rapidly, outpacing governance mechanisms designed to ensure safety and accountability.
+
+**Key data:**
+- Major tech firms projected to spend ~$405 billion building AI infrastructure in 2025
+- Four largest tech providers may invest "$650 billion more" in 2026
+- Sequoia Capital identified "a $600 billion gap between AI infrastructure spending and AI earnings" — intense pressure to monetize capabilities quickly
+- 63% of surveyed organizations lack AI governance policies (IBM research)
+
+**Key claims:**
+1. Rapid deployment velocity creates systemic risk when low-probability failures scale across millions of users
+2. Regulatory timelines (years) cannot match AI release cycles (weeks to hours)
+3. Organizations face reputational, legal, and operational risks from inadequate governance
+4. Strong governance functions as competitive advantage, not merely compliance burden
+
+**Proposed organizational governance framework:**
+- Risk assessment before deployment
+- Design-integrated risk mitigation
+- Auditability and accountability pathways
+- Monitoring and incident response plans
+- Data protection measures
+
+## Agent Notes
+
+**Why this matters:** The investment data ($405B infrastructure in 2025, $650B planned 2026, $600B Sequoia gap) quantifies the scale mismatch between capability investment and governance investment. This is the structural dynamic that enables all four overshoot mechanisms: the financial pressure to monetize creates the competitive adoption cycle, which drives the "follow or die" dynamic, which drives overshoot.
+
+**What surprised me:** 63% of organizations lack AI governance policies despite all the regulatory activity (EU AI Act, NIST RMF, etc.) — much higher than I expected. This confirms the governance deficit is not theoretical but empirically widespread.
+
+**What I expected but didn't find:** Comparative data on governance investment vs. capability investment (would need something like "safety budgets as % of capability R&D"). The piece has capability investment data but not governance investment data.
+
+**KB connections:**
+- [[technology advances exponentially but coordination mechanisms evolve linearly creating a widening gap]] — the quantitative version: $1.05T in AI infrastructure vs. governance that evolves on regulatory timelines
+- [[safe AI development requires building alignment mechanisms before scaling capability]] — the $600B Sequoia gap is direct evidence this sequencing rule is being violated
+- [[voluntary safety pledges cannot survive competitive pressure]] — the $600B monetization gap IS the competitive pressure mechanism
+
+**Extraction hints:**
+- Not much to extract as new claims — this largely confirms existing KB claims with new data. Most valuable as evidence enrichment.
+- Could update [[technology advances exponentially but coordination mechanisms evolve linearly]] with the quantitative data: $1.05T infrastructure, $600B Sequoia gap, 63% lacking governance policies.
+- The "strong governance as competitive advantage" claim is potentially extractable if there's evidence behind it — but the article asserts it rather than demonstrates it.
+
+**Context:** Strategy International is a UK-based think tank. Publication is timely (March 11, 2026). Standard quality, not peer-reviewed.
+
+## Curator Notes
+
+PRIMARY CONNECTION: [[technology advances exponentially but coordination mechanisms evolve linearly creating a widening gap]]
+
+WHY ARCHIVED: Provides quantitative scale data ($405B/$650B investment, $600B Sequoia gap, 63% governance deficit) that gives concrete numbers to the abstract coordination gap. Most useful as evidence enrichment for existing claims rather than new claim extraction.
+
+EXTRACTION HINT: Use primarily as evidence enrichment for existing claims about investment-governance mismatch. Note the $600B Sequoia gap as the specific monetization pressure mechanism.

inbox/queue/2026-03-18-cfr-how-2026-decides-ai-future-governance.md

@@ -0,0 +1,63 @@
+---
+type: source
+title: "How 2026 Could Decide the Future of Artificial Intelligence"
+author: "Council on Foreign Relations (multiple fellows)"
+url: https://www.cfr.org/articles/how-2026-could-decide-future-artificial-intelligence
+date: 2026-03-18
+domain: ai-alignment
+secondary_domains: []
+format: article
+status: unprocessed
+priority: medium
+tags: [governance, international-coordination, EU-AI-Act, enforcement, geopolitics, 2026-inflection]
+---
+
+## Content
+
+**Core framing:** 2026 represents a pivotal shift from AI speculation to operational reality — regulatory frameworks colliding with actual deployment at scale.
+
+**Key governance claims from six CFR fellows:**
+
+1. **Kat Duffy:** "Truly operationalizing AI governance will be the sticky wicket of 2026." Implementation, not design, is the challenge.
+
+2. **Vinh Nguyen:** Three pillars for trustworthy AI deployment: threat intelligence platforms monitoring AI use; continuous validation of machine identities; governed channels for AI tools with mandatory production code reviews.
+
+3. **Michael Horowitz:** US must engage in "standard-setting bodies" to counter China's AI governance influence. Notes: "large-scale binding international agreements on AI governance are unlikely in 2026."
+
+**Enforcement mechanisms noted:**
+- EU AI Act: penalties up to €35 million or 7% of global turnover
+- China's amended Cybersecurity Law emphasizing state oversight
+- U.S. state-level rules taking effect across 2026
+- "One Big Beautiful Bill Act" appropriating billions for Pentagon AI priorities
+
+**Autonomous AI systems raising questions:** Legal accountability and responsibility assignment unresolved for AI decisions with no clear human author.
+
+**Diverging governance philosophies:** Democracies vs. authoritarian systems creating different AI governance approaches and potential strategic advantages.
+
+## Agent Notes
+
+**Why this matters:** Confirms the disconfirmation search result: large-scale binding international agreements are "unlikely in 2026" per Horowitz. The governance that IS happening is enforcement of existing frameworks (EU AI Act), US/China strategic divergence, and bilateral procurement negotiations — not the multilateral coordination that would actually address the structural race dynamics. The "operationalization problem" (governance designed, not yet implemented) is the key gap.
+
+**What surprised me:** Michael Horowitz explicitly saying binding international agreements are unlikely in 2026 — from a CFR fellow, this is a notable concession about the limits of international governance coordination. Most governance commentary is more optimistic.
+
+**What I expected but didn't find:** Any specific mechanism for how autonomous AI accountability will be resolved. The article identifies it as an unresolved problem but doesn't propose solutions.
+
+**KB connections:**
+- [[AI development is a critical juncture in institutional history where the mismatch between capabilities and governance creates a window for transformation]] — this CFR piece is the policy establishment's view of where that window stands
+- [[technology advances exponentially but coordination mechanisms evolve linearly]] — the "operationalization problem" is a specific instance: governance designed but implementation lagging deployment
+- [[multipolar failure from competing aligned AI systems may pose greater existential risk]] — US/China governance divergence is exactly the multipolar dynamic that creates interaction risks
+
+**Extraction hints:**
+- Not much new to extract — mainly confirmation of existing claims with policy establishment framing.
+- The "binding international agreements unlikely in 2026" claim from Horowitz is quotable for updating existing governance claims.
+- The autonomous AI accountability gap (no mechanism for responsibility when AI makes decisions with no clear human author) could be a claim candidate: "current legal accountability frameworks cannot assign responsibility for autonomous AI decisions because they require a human decision-maker as the legal subject"
+
+**Context:** CFR is mainstream US foreign policy establishment. Six fellows contributing = diverse perspectives. Published March 2026.
+
+## Curator Notes
+
+PRIMARY CONNECTION: [[AI development is a critical juncture in institutional history where the mismatch between capabilities and governance creates a window for transformation]]
+
+WHY ARCHIVED: Provides establishment policy view on 2026 AI governance landscape. Most valuable for confirming the international coordination failure (binding agreements unlikely). The legal accountability gap for autonomous AI decisions may be worth extracting.
+
+EXTRACTION HINT: Use for evidence enrichment on coordination gap claims. The legal accountability claim ("autonomous AI, no human author") may be worth extracting if not already in KB.

inbox/queue/2026-03-18-hks-governance-by-procurement-bilateral.md

@@ -0,0 +1,55 @@
+---
+type: source
+title: "Governance by Procurement: How AI Rights Became a Bilateral Negotiation"
+author: "Harvard Kennedy School — Carr-Ryan Center for Human Rights"
+url: https://www.hks.harvard.edu/centers/carr-ryan/our-work/carr-ryan-commentary/governance-procurement-how-ai-rights-became
+date: 2026-03-18
+domain: ai-alignment
+secondary_domains: []
+format: article
+status: unprocessed
+priority: high
+tags: [governance, procurement, bilateral-negotiation, international-coordination, anthropic, DoD, correction-failure, transparency]
+---
+
+## Content
+
+**Core argument:** The most consequential AI governance decisions are being made through private contracts between governments and technology companies, not through multilateral democratic processes. "The most consequential human rights questions in AI are being decided in bilateral negotiations between governments and technology companies. Most of the world is not in the room."
+
+**The mechanism:** International human rights protections now depend on individual corporate leaders' ethical choices — governance conducted "without transparency, without public accountability, and without remedy mechanisms for those affected."
+
+**Centerpiece example:** A 2026 confrontation where the Department of War (formerly Defense) threatened to blacklist Anthropic unless it removed safeguards against mass surveillance and autonomous weapons. Anthropic refused publicly. Pentagon retaliated. This illustrates how critical protections depend on individual corporate decisions, not binding international frameworks.
+
+**Proposed corrections (multilateral):**
+- Technical standards through the International Telecommunication Union (ITU)
+- Global Digital Compact grounding AI governance in human rights law
+- ISO/IEC standards for AI management systems
+- International AI oversight body modeled after nuclear energy regulation
+
+## Agent Notes
+
+**Why this matters:** This is a direct confirmation of the keystone belief disconfirmation search. The question was: "are governance mechanisms keeping pace with AI capabilities?" The HKS analysis says NO — and more precisely, the governance that IS happening is bilateral, opaque, and structurally captured by the power asymmetry between governments and labs. The DoD/Anthropic confrontation is a concrete example of the government as coordination-BREAKER (threatening to penalize safety constraints), not correction mechanism.
+
+**What surprised me:** The DoD reportedly threatened to BLACKLIST Anthropic for maintaining safety safeguards — this is a direct government role as alignment-degrader. This is a new development beyond what was in our KB. The existing claim [[government designation of safety-conscious AI labs as supply chain risks inverts the regulatory dynamic]] may need updating with this specific episode.
+
+**What I expected but didn't find:** Evidence that the proposed multilateral alternatives (ITU, Global Digital Compact) are advancing at pace with the bilateral negotiation pattern. The article proposes these but doesn't assess their current momentum.
+
+**KB connections:**
+- [[government designation of safety-conscious AI labs as supply chain risks inverts the regulatory dynamic by penalizing safety constraints rather than enforcing them]] — the DoD/Anthropic episode is a specific instance of this pattern
+- [[voluntary safety pledges cannot survive competitive pressure because unilateral commitments are structurally punished when competitors advance without equivalent constraints]] — the Anthropic case shows government adding to competitive pressure
+- [[AI development is a critical juncture in institutional history where the mismatch between capabilities and governance creates a window for transformation]] — this piece shows how the window is being used (and misused)
+
+**Extraction hints:**
+- Claim candidate: "bilateral government-tech company negotiations are the de facto AI governance mechanism in 2026, bypassing multilateral frameworks and making human rights protections contingent on individual corporate decisions"
+- The DoD/Anthropic confrontation may need careful claim scoping — it's one episode. The broader pattern of bilateral negotiation is the extractable claim.
+- Update consideration: [[government designation of safety-conscious AI labs as supply chain risks]] — this episode should be added as additional evidence.
+
+**Context:** HKS Carr-Ryan Center for Human Rights is highly credible. The DoD/Anthropic episode is striking and should be verified — this could be the most significant development in the AI governance space in months.
+
+## Curator Notes
+
+PRIMARY CONNECTION: [[government designation of safety-conscious AI labs as supply chain risks inverts the regulatory dynamic by penalizing safety constraints rather than enforcing them]]
+
+WHY ARCHIVED: Confirms that government as correction mechanism is FAILING — more specifically, government is sometimes functioning as a coordination-BREAKER. This directly addresses the disconfirmation search for B1 (keystone belief). The DoD/Anthropic episode is the most concrete governance failure example since Anthropic RSP rollback.
+
+EXTRACTION HINT: Extract the bilateral negotiation claim with specific evidence. Also flag for enrichment of existing claim about government-as-supply-chain-risk with the DoD confrontation example.

inbox/queue/2026-03-18-nbcnews-aiuc-ai-insurance-safety-market-mechanism.md

@@ -0,0 +1,67 @@
+---
+type: source
+title: "Insurance Companies Are Trying to Make AI Safer"
+author: "NBC News Technology Desk"
+url: https://www.nbcnews.com/tech/tech-news/insurance-companies-are-trying-to-make-ai-safer-rcna243834
+date: 2026-03-18
+domain: ai-alignment
+secondary_domains: [internet-finance]
+format: article
+status: unprocessed
+priority: high
+tags: [insurance, market-mechanism, AIUC, safety-certification, skin-in-the-game, correction-mechanism, accountability]
+flagged_for_rio: ["Market-based AI safety mechanism with insurance economics — Rio should evaluate whether this has properties analogous to prediction markets for surfacing true risk probabilities"]
+---
+
+## Content
+
+Main claim: Insurance companies are positioning as market-based regulators of AI safety, arguing they can incentivize safer AI practices by making coverage contingent on risk mitigation — without waiting for government oversight.
+
+**AIUC (Artificial Intelligence Underwriting Company):** Insurance startup developing industry standards. In July 2025, launched AIUC-1 — "the world's first certification for AI agents." Standard covers six pillars:
+1. Security
+2. Safety
+3. Reliability
+4. Data and privacy
+5. Accountability
+6. Societal risks
+
+Michael von Gablenz (Munich Re): "Insurance has played a major role in [safety improvements], and I believe insurance can play the same role for AI."
+
+**Historical precedent cited:**
+- Benjamin Franklin's 1700s fire insurance company → precursor to modern building codes (required safety standards for coverage)
+- Seatbelt adoption → driven by insurance premium incentives, not government mandate alone
+
+**Market mechanisms:**
+1. Financial incentives: Insurers profit by accurately pricing risk and preventing claims → incentivize AI developers to make safer products
+2. Certification requirements: Safety standards required before coverage → creates pre-market safety pressure
+3. Claims data collection: Insurers track losses → identify which practices actually prevent harm → share findings with developers (information aggregation)
+
+**Market size:** AI insurance market projected at ~$4.7B in premiums by 2032.
+
+## Agent Notes
+
+**Why this matters:** First evidence of a market-based correction mechanism with genuine skin-in-the-game properties for AI safety. Insurance is uniquely positioned: (a) it has financial incentives to accurately measure outcomes (unlike self-reporting), (b) it creates pre-market pressure through certification requirements, (c) it has historical precedent as a correction mechanism in other high-stakes domains. This is the closest analog to the prediction markets approach Rio would recognize.
+
+**What surprised me:** The AIUC-1 certification exists and was launched in July 2025 — this is more developed than I expected. Also surprising: the historical precedent (Franklin's fire insurance → building codes) suggests insurance has successfully driven safety standards before regulatory frameworks existed. This is a genuine market-before-government correction pathway.
+
+**What I expected but didn't find:** Evidence that insurance premiums are actually differential enough to incentivize safety investment (vs. just covering the risk). Also missing: how AIUC-1 certification interacts with the perception gap problem — insurers need accurate outcome data, but the perception gap (METR RCT: 39-point gap) means self-reported incident data is unreliable.
+
+**KB connections:**
+- [[the alignment tax creates a structural race to the bottom because safety training costs capability and rational competitors skip it]] — insurance could internalize the alignment tax
+- [[voluntary safety pledges cannot survive competitive pressure]] — insurance creates enforceable (not just voluntary) standards
+- [[economic forces push humans out of every cognitive loop]] — this mechanism pushes back through premium incentives
+
+**Extraction hints:**
+- Claim candidate: "AI liability insurance is emerging as a market-based correction mechanism for automation overshoot because it creates financial incentives for safety measurement that don't depend on government coordination or voluntary commitments"
+- Note the critical limitation: insurance requires accurate outcome measurement, which the perception gap (METR RCT) undermines. The claim needs this scoping.
+- The historical precedent (fire insurance → building codes; seatbelts + insurance) is separately extractable as evidence that insurance has successfully driven safety standards before regulatory frameworks.
+
+**Context:** NBC News tech desk — general interest, not technical. Munich Re is the world's largest reinsurer and deeply credible. AIUC is early-stage.
+
+## Curator Notes
+
+PRIMARY CONNECTION: [[the alignment tax creates a structural race to the bottom because safety training costs capability and rational competitors skip it]] — insurance inverts this by making safety non-adoption costly
+
+WHY ARCHIVED: First identified correction mechanism with genuine skin-in-the-game properties. Also flagged for Rio due to mechanism design relevance.
+
+EXTRACTION HINT: Extract the insurance-as-correction-mechanism claim with explicit scoping about the measurement dependency. The historical precedent deserves a separate extraction.