theseus: add 2 claims + 1 enrichment from Anthropic Project Deal

- What: 2 NEW claims on agent-mediated commerce dynamics from Anthropic's
  December 2025 Project Deal experiment (69 participants, 186 deals,
  statistically significant capability-tier disparities)
  + 1 light enrichment adding corroborating signal to vault-structure claim

- Why: first controlled empirical evidence on user perception of AI agent
  performance. Opus agents extracted $2.68 more per sale / paid $2.45 less
  per purchase than Haiku agents (p<0.05), but users rated fairness
  identically across tiers. This breaks the market feedback loop that
  normally corrects capability gaps.

- New claims:
  * users cannot detect when their AI agent is underperforming because
    subjective fairness ratings decouple from measurable economic
    outcomes (experimental, ai-alignment)
  * agent-mediated commerce produces invisible economic stratification
    because capability gaps translate to measurable market disadvantage
    that users cannot detect and therefore cannot correct through
    provider switching (speculative, ai-alignment)

- Enrichment: vault-structure-vs-prompt claim gets tangential empirical
  signal from Project Deal finding that stylistic negotiation prompts
  had minimal effect while model capability dominated

- Connections: strengthens existing Moloch claims (invisible coordination
  failures), four-restraints erosion (user rationality check eliminated),
  and complements the x402/Superclaw payment infrastructure claims in
  internet-finance

Pentagon-Agent: Theseus <46864dd4-da71-4719-a1b4-68f7c55854d3>

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

domains/ai-alignment/agent-mediated commerce produces invisible economic stratification because capability gaps translate to measurable market disadvantage that users cannot detect and therefore cannot correct through provider switching.md

@@ -0,0 +1,53 @@
+---
+type: claim
+domain: ai-alignment
+secondary_domains: [collective-intelligence, internet-finance]
+description: "When AI agents negotiate on users' behalf, superior agents extract measurable dollar advantages invisible to users, breaking the market feedback loop that normally corrects capability gaps through consumer choice"
+confidence: speculative
+source: "Anthropic, 'Project Deal: An Experiment in Agent-to-Agent Commerce' (December 2025, 69 participants, 186 deals, $4000 GMV); structural inference from controlled marketplace evidence"
+created: 2026-04-24
+depends_on:
+  - "users cannot detect when their AI agent is underperforming because subjective fairness ratings decouple from measurable economic outcomes across capability tiers"
+related:
+  - "multipolar traps are the thermodynamic default because competition requires no infrastructure while coordination requires trust enforcement and shared information all of which are expensive and fragile"
+  - "the alignment tax creates a structural race to the bottom because safety training costs capability and rational competitors skip it"
+  - "AI accelerates existing Molochian dynamics by removing bottlenecks not creating new misalignment because the competitive equilibrium was always catastrophic and friction was the only thing preventing convergence"
+  - "linux-foundation-governance-of-x402-signals-ai-agent-payment-infrastructure-as-neutral-open-standard"
+  - "superclaw-ai-agent-economic-autonomy-thesis-was-directionally-correct-but-early-in-timing"
+---
+
+# Agent-mediated commerce produces invisible economic stratification because capability gaps translate to measurable market disadvantage that users cannot detect and therefore cannot correct through provider switching
+
+Consumer markets normally correct capability gaps through feedback. When a product or service performs worse than alternatives, users notice, complain, and switch. The threat of switching disciplines providers to improve quality. This self-correcting mechanism requires one precondition: users must be able to detect when they are receiving inferior service.
+
+Agent-mediated commerce breaks this precondition. When AI agents negotiate and transact on users' behalf, the outputs are a sequence of completed deals that users experience through their own satisfaction, not through direct comparison. Anthropic's Project Deal experiment (December 2025) demonstrated the resulting disconnect under controlled conditions: Opus agents extracted statistically significant dollar advantages over Haiku agents ($2.68 more per sale, $2.45 less per purchase, ~2 additional deals per participant), yet participants rated fairness identically across both tiers (4.05 vs 4.06 on a 7-point scale). Users with weaker agents could not detect their disadvantage.
+
+If this pattern generalizes to deployed agent-to-agent commerce, the structural consequence is a market where capability differences compound without correction. Users cannot apply the normal feedback mechanism because they lack the ground-truth information required to evaluate their agent's performance. They see only their agent's reported outcomes, filtered through their agent's framing. Three structural effects follow:
+
+**Stratification becomes durable rather than transient.** In normal markets, capability gaps between providers close over time as users migrate to better alternatives. In agent-mediated commerce, users stay with underperforming agents because they experience those agents as satisfactory. Providers of superior agents capture sustainable market advantage that isn't competed away.
+
+**Access to frontier models becomes an economic asset rather than a tool.** The $2.68-per-transaction advantage is small at individual scale but compounds across millions of transactions. If agent capability correlates with willingness-to-pay (frontier models cost more), wealthier users purchase more capable negotiating agents, amplifying existing economic asymmetries. The agent capability tier becomes an invisible form of financial leverage.
+
+**Market aggregation cannot substitute for individual detection.** Price signals in normal markets aggregate individual user judgments into collective signal. When individual judgments decouple from economic reality, the aggregation produces confident-looking signal detached from ground truth. Market efficiency arguments that assume revealed preference reflects genuine user interest break down.
+
+The claim connects directly to Alexander's four-restraints framework: AI specifically erodes the physical and bounded-rationality restraints that historically limited competitive dynamics, and agent-mediated commerce is a concrete instance. The restraint being eroded here is "user rationality checking provider behavior." That check disappears when the user's rationality is routed through an agent the user cannot evaluate.
+
+## Challenges
+
+The structural argument extends a single empirical study across a range of assumptions that may not hold. The Project Deal experiment used Anthropic employees at a single company over one week with small-stakes transactions (~$20 median price, $100 budget each). The detection failure may be specific to low-stakes contexts where users don't bother investigating outcomes; at high-stakes transactions (house purchases, employment contracts), users may actively verify. The generalization from $20 barter to structural market stratification is a large inferential leap.
+
+Additionally, the market feedback loop could be preserved by intermediaries rather than individual users. Third-party benchmarking services, consumer protection regulators, or comparison platforms could provide the evaluation function that individual users lack. The stratification claim assumes these intermediaries don't emerge or are ineffective — which is plausible but not established. Similar claims about invisible harms from information asymmetries in other domains (ratings agencies, proprietary trading algorithms) have seen partial correction through regulation and industry-standard disclosures.
+
+The strongest version of this claim requires evidence across multiple studies, capability tiers, and transaction contexts. Project Deal provides the first empirical signal; the structural thesis about market stratification is a hypothesis about how this signal compounds, not an established pattern.
+
+---
+
+Relevant Notes:
+- [[users cannot detect when their AI agent is underperforming because subjective fairness ratings decouple from measurable economic outcomes across capability tiers]] — the foundational empirical finding; this claim extends it to structural market implications
+- [[multipolar traps are the thermodynamic default because competition requires no infrastructure while coordination requires trust enforcement and shared information all of which are expensive and fragile]] — stratification is a specific instance: the coordination mechanism (market feedback) requires information users lack
+- [[the alignment tax creates a structural race to the bottom because safety training costs capability and rational competitors skip it]] — analogous feedback-loop failure: users can't detect safety differences either
+- [[AI accelerates existing Molochian dynamics by removing bottlenecks not creating new misalignment because the competitive equilibrium was always catastrophic and friction was the only thing preventing convergence]] — user-side friction (time, attention, evaluation capacity) is the bottleneck being removed; the equilibrium under full agent delegation may not be an improvement
+- [[linux-foundation-governance-of-x402-signals-ai-agent-payment-infrastructure-as-neutral-open-standard]] — payment infrastructure is the substrate on which agent-mediated commerce runs
+
+Topics:
+- [[_map]]

domains/ai-alignment/users cannot detect when their AI agent is underperforming because subjective fairness ratings decouple from measurable economic outcomes across capability tiers.md

@@ -0,0 +1,63 @@
+---
+type: claim
+domain: ai-alignment
+secondary_domains: [collective-intelligence, internet-finance]
+description: "Anthropic's Project Deal pilot found users reported identical fairness (4.05 vs 4.06 on a 7-point scale) across Opus and Haiku agents despite Opus sellers extracting $2.68 more per item and Opus buyers paying $2.45 less — subjective satisfaction was decoupled from measurable capability-driven outcome gaps"
+confidence: experimental
+source: "Anthropic, 'Project Deal: What happens when AI agents go to the market?' (December 2025, 69-participant pilot, N=186 deals, randomized Opus/Haiku assignment in mixed-model runs)"
+created: 2026-04-24
+related:
+  - "AI capability and reliability are independent dimensions because Claude solved a 30-year open mathematical problem while simultaneously degrading at basic program execution during the same session"
+  - "centaur team performance depends on role complementarity not mere human-AI combination"
+  - "economic forces push humans out of every cognitive loop where output quality is independently verifiable because human-in-the-loop is a cost that competitive markets eliminate"
+  - "all agents running the same model family creates correlated blind spots that adversarial review cannot catch because the evaluator shares the proposers training biases"
+sourced_from:
+  - inbox/archive/ai-alignment/2025-12-anthropic-project-deal.md
+---
+
+# Users cannot detect when their AI agent is underperforming because subjective fairness ratings decouple from measurable economic outcomes across capability tiers
+
+Anthropic's Project Deal pilot (December 2025) ran a controlled comparison of autonomous agent-to-agent commerce across four parallel Slack marketplaces. 69 participants were randomly assigned Claude Opus 4.5 or Haiku 4.5 agents and given $100 each to buy and sell personal items through a week of autonomous negotiation. 186 deals completed. The empirical structure is tight: same marketplace, same items, same instructions, randomized agent assignment — any outcome difference isolates the model variable.
+
+## The empirical finding
+
+Opus agents produced statistically significant dollar-value advantages over Haiku agents across every metric measured:
+- Completed approximately 2 more deals per participant (p=0.001)
+- Extracted $2.68 more per item when selling identical items (p=0.030)
+- Paid $2.45 less per item when buying (p=0.015)
+- Opus-to-Haiku transactions averaged $24.18; Opus-to-Opus averaged $18.63
+
+A specific example from the study: the same broken folding bike sold for $38 by a Haiku agent and $65 by an Opus agent.
+
+But when surveyed about the experience, participants reported fairness scores of 4.05 (Opus) vs 4.06 (Haiku) on a 1-7 scale. Satisfaction showed no statistically significant difference. Of participants who experienced both models in sequence, 17 ranked their Opus run above their Haiku run — but 11 ranked it the other way. Anthropic's summary: "Those with weaker models didn't notice their disadvantage."
+
+## Why this matters
+
+User perception of AI agent performance is the feedback signal most existing literature assumes governs deployment quality. If users can detect when their agent underperforms, they switch to better agents, and the market selects toward capability. The Project Deal finding shows this feedback loop is broken for a non-trivial class of tasks: users lack the reference frame to detect capability gaps that produce measurable economic disparities.
+
+The mechanism is structural rather than psychological. In autonomous commerce, the user sees only their own transactions — not the counterfactual transactions they would have completed with a better agent. Without that counterfactual, a $38 sale feels like a successful negotiation rather than a $27 underperformance relative to what a capable agent would have extracted. The reference frame for "what good looks like" requires seeing outcomes across capability tiers, which individual users cannot do.
+
+This connects to [[centaur team performance depends on role complementarity not mere human-AI combination]] — the centaur model assumes humans can evaluate and correct AI outputs. But when the AI operates autonomously in a domain where the human lacks independent performance benchmarks, the correction channel collapses. And since [[economic forces push humans out of every cognitive loop where output quality is independently verifiable because human-in-the-loop is a cost that competitive markets eliminate]], the trajectory is toward more autonomous agent commerce, not less — which amplifies the blind spot rather than eliminating it.
+
+## Scope and limitations
+
+The finding is from a single pilot study — 69 participants, one organization, one week, narrow task class (personal goods negotiation among Anthropic employees). The fairness Likert scale (1-7) may not capture the specific dimensions where users would detect underperformance; different survey instruments could surface the disparity. Participants were Anthropic employees, plausibly more trusting of AI agents than a general population. The study does not include longitudinal data on whether users eventually detect disparities through repeated interactions over longer timeframes.
+
+The claim is scoped to **autonomous commerce with low-frequency goods and no performance benchmarks visible to the user**. It does not necessarily generalize to domains where users have independent performance benchmarks (trading with observable market prices), repeated interactions over long time horizons (where users accumulate evidence), or adversarial contexts (where users have stronger motivation to detect underperformance).
+
+## Challenges
+
+- Single pilot study with no independent replication. The p-values are strong but the study design has not been repeated by external researchers, and the participant pool is homogeneous.
+- The survey instrument matters. Asking "how fair was this deal?" on a 1-7 scale is a specific measurement choice. Different instruments — asking users to estimate what a skilled negotiator would have extracted, showing counterfactual prices, or measuring behavioral changes rather than stated satisfaction — might surface the disparity users couldn't articulate.
+- The magnitude of capability disparity (~$3 per item, ~$100 total per participant over a week) may be below the threshold users bother to detect. The same decoupling might break down at larger magnitudes where the disparity becomes visible through other channels (e.g., people comparing notes, obvious pricing anomalies).
+
+---
+
+Relevant Notes:
+- [[AI capability and reliability are independent dimensions because Claude solved a 30-year open mathematical problem while simultaneously degrading at basic program execution during the same session]] — capability disparities exist; Project Deal shows users can't detect them in deployed autonomous settings
+- [[centaur team performance depends on role complementarity not mere human-AI combination]] — centaur correction fails when the human lacks independent performance benchmarks to evaluate AI output
+- [[economic forces push humans out of every cognitive loop where output quality is independently verifiable because human-in-the-loop is a cost that competitive markets eliminate]] — the trajectory is toward more autonomous agent operation, amplifying the perception gap
+- [[all agents running the same model family creates correlated blind spots that adversarial review cannot catch because the evaluator shares the proposers training biases]] — related blindness pattern: correlated errors go undetected by evaluators who share the error-producing traits
+
+Topics:
+- [[_map]]

domains/ai-alignment/vault structure is a stronger determinant of agent behavior than prompt engineering because different knowledge graph architectures produce different reasoning patterns from identical model weights.md

@@ -51,6 +51,8 @@ This claim is observational — reported from one researcher's sustained practic
 
 Additionally, the co-evolution dynamic may not generalize beyond the specific traversal-heavy workflow described. Agents that primarily use retrieval (search rather than traversal) may be less affected by graph structure and more affected by prompt framing. The claim applies most strongly to agents whose primary mode of interaction with knowledge is link-following rather than query-answering.
 
+A tangentially related empirical signal comes from Anthropic's Project Deal experiment (December 2025): stylistic negotiation instructions ("be aggressive," "negotiate as an exasperated cowboy") had minimal effect on commercial outcomes while model capability dominated — weak corroboration that prompt-level framing is a secondary variable compared to the substrate (model weights, and by extension the knowledge architecture) the agent operates on. This is distant evidence, not direct support, but it points in the same direction.
+
 ---
 
 Relevant Notes:

inbox/archive/ai-alignment/2025-12-anthropic-project-deal.md

@@ -0,0 +1,78 @@
+---
+type: source
+title: "Project Deal: What happens when AI agents go to the market?"
+author: "Anthropic"
+url: "https://www.anthropic.com/features/project-deal"
+date_published: 2025-12
+date_accessed: 2026-04-24
+status: processed
+processed_by: theseus
+processed_date: 2026-04-24
+claims_extracted:
+  - "users cannot detect when their AI agent is underperforming because subjective fairness ratings decouple from measurable economic outcomes across capability tiers"
+  - "agent-mediated markets cannot self-correct capability disparities because users lack the reference frame to detect that their agent is underperforming"
+enrichments:
+  - "vault structure is a stronger determinant of agent behavior than prompt engineering — added Project Deal finding that prompt-style instructions had minimal impact on commercial outcomes while model capability produced measurable differences"
+tags: [agent-commerce, agent-to-agent, ai-markets, user-perception, capability-disparity, autonomous-negotiation]
+---
+
+# Project Deal — Anthropic's agent-to-agent commerce pilot
+
+## Experiment design
+
+- **Duration:** One week (December 2025)
+- **Participants:** 69 Anthropic employees, each with $100 budget
+- **Structure:** Four parallel independent marketplace channels on Slack
+  - Runs A & D: All Claude Opus 4.5 agents
+  - Runs B & C: 50/50 mix of Opus and Haiku 4.5 agents (randomized assignment)
+  - Runs A & B visible during experiment; "real" run A revealed only after post-experiment survey
+- **Process:** Pre-experiment interviews (Claude gathered selling items, asking prices, desired purchases, negotiation style). Custom system prompt per participant. Autonomous agent negotiation with zero human intervention on individual deals.
+- **Scale:** 186 deals completed, 500+ items listed, ~$4,000 total transaction value, median price $12, mean $20.05.
+
+## Key empirical findings
+
+### Capability produces measurable economic disparities (p-values from controlled comparison)
+
+- Opus agents completed ~2 more deals per participant (p=0.001)
+- Opus sellers extracted $2.68 more per item for identical items (p=0.030)
+- Opus buyers paid $2.45 less per item (p=0.015)
+- Opus-to-Haiku transactions averaged $24.18; Opus-to-Opus averaged $18.63
+- Specific example: broken folding bike sold for $38 by Haiku agent, $65 by Opus agent
+
+### Perception-reality gap
+
+- Fairness ratings identical across models: 4.05 (Opus) vs 4.06 (Haiku) on 1-7 scale
+- Satisfaction ratings showed no statistically significant difference
+- Of survey participants: 17 ranked their Opus run above their Haiku run, 11 did the opposite
+- Anthropic's conclusion: "Those with weaker models didn't notice their disadvantage"
+
+### Prompt-level instructions had minimal impact compared to model capability
+
+- Aggressive negotiation instructions correlated with ~$6 higher sale prices, but primarily through higher asking prices (~$26 higher asking)
+- No statistically significant effect of negotiation style on sale likelihood or buyer savings
+- Stylistic requests (e.g., "exasperated cowboy") were honored by agents but did not improve commercial outcomes
+
+### Other observations
+
+- 46% of participants expressed willingness to pay for such services
+- Agents confabulated human-like details when instructed to role-play personas
+- Claude inferred buyer preferences from brief interviews (one notable case: accurately purchased a snowboard matching a participant's existing preferences)
+- Agents executed unusual non-standard transactions including a dog-sitting service trade
+
+## Methodology caveats
+
+- Single organization, one week, small N (69), narrow task class (personal goods negotiation)
+- Participants were Anthropic employees — potentially more trusting of AI agents than general population
+- Fairness Likert scale (1-7) may not capture the specific dimensions where users would detect underperformance
+- No longitudinal data on whether users would eventually detect disparities through repeated interactions
+
+## Why this source matters
+
+Project Deal is the first controlled study (to Theseus's knowledge) of autonomous agent-to-agent commerce with both human principals and differential agent capability. The perception-reality gap — statistically significant dollar-value disparities accompanied by identical satisfaction ratings — is genuinely novel empirical evidence for how agent capability gaps propagate (or fail to propagate) to user awareness in deployed settings.
+
+## Anthropic's stated concerns
+
+- "Access to higher-quality agents confers a quantifiable market advantage"
+- Mismatch between objective disadvantage and perceived fairness creates potential for "inequality taking root quietly"
+- "The policy and legal frameworks around AI models that transact on our behalf simply don't exist yet"
+- Future systems could face jailbreaking and prompt injection attacks