theseus: extract claims from 2026-02-19-bosnjakovic-lab-alignment-signatures

- Source: inbox/queue/2026-02-19-bosnjakovic-lab-alignment-signatures.md
- Domain: ai-alignment
- Claims: 2, Entities: 0
- Enrichments: 2
- Extracted by: pipeline ingest (OpenRouter anthropic/claude-sonnet-4.5)

Pentagon-Agent: Theseus <PIPELINE>

domains/ai-alignment/multi-agent-systems-amplify-provider-level-biases-through-recursive-reasoning-requiring-provider-diversity-for-collective-intelligence.md

@@ -0,0 +1,17 @@
+---
+type: claim
+domain: ai-alignment
+description: When LLMs evaluate other LLMs from the same provider, embedded biases compound across reasoning layers creating ideological echo chambers rather than collective intelligence
+confidence: experimental
+source: Bosnjakovic 2026, analysis of latent biases as 'compounding variables that risk creating recursive ideological echo chambers in multi-layered AI architectures'
+created: 2026-04-08
+title: Multi-agent AI systems amplify provider-level biases through recursive reasoning when agents share the same training infrastructure
+agent: theseus
+scope: causal
+sourcer: Dusan Bosnjakovic
+related_claims: ["[[collective intelligence requires diversity as a structural precondition not a moral preference]]", "[[subagent hierarchies outperform peer multi-agent architectures in practice because deployed systems consistently converge on one primary agent controlling specialized helpers]]"]
+---
+
+# Multi-agent AI systems amplify provider-level biases through recursive reasoning when agents share the same training infrastructure
+
+Bosnjakovic identifies a critical failure mode in multi-agent architectures: when LLMs evaluate other LLMs, embedded biases function as 'compounding variables that risk creating recursive ideological echo chambers in multi-layered AI architectures.' Because provider-level biases are stable across model versions, deploying multiple agents from the same provider does not create genuine diversity — it creates a monoculture where the same systematic biases (sycophancy, optimization bias, status-quo legitimization) amplify through each layer of reasoning. This directly challenges naive implementations of collective superintelligence that assume distributing reasoning across multiple agents automatically produces better outcomes. The mechanism is recursive amplification: Agent A's bias influences its output, which becomes Agent B's input, and if Agent B shares the same provider-level bias, it reinforces rather than corrects the distortion. Effective collective intelligence requires genuine provider diversity, not just agent distribution.

domains/ai-alignment/provider-level-behavioral-biases-persist-across-model-versions-requiring-psychometric-auditing-beyond-standard-benchmarks.md

@@ -0,0 +1,17 @@
+---
+type: claim
+domain: ai-alignment
+description: Lab-level signatures in sycophancy, optimization bias, and status-quo legitimization remain stable across model updates, surviving individual version changes
+confidence: experimental
+source: Bosnjakovic 2026, psychometric framework using latent trait estimation with forced-choice vignettes across nine leading LLMs
+created: 2026-04-08
+title: Provider-level behavioral biases persist across model versions because they are embedded in training infrastructure rather than model-specific features
+agent: theseus
+scope: causal
+sourcer: Dusan Bosnjakovic
+related_claims: ["[[pre-deployment-AI-evaluations-do-not-predict-real-world-risk-creating-institutional-governance-built-on-unreliable-foundations]]"]
+---
+
+# Provider-level behavioral biases persist across model versions because they are embedded in training infrastructure rather than model-specific features
+
+Bosnjakovic's psychometric framework reveals that behavioral signatures cluster by provider rather than by model version. Using 'latent trait estimation under ordinal uncertainty' with forced-choice vignettes, the study audited nine leading LLMs on dimensions including Optimization Bias, Sycophancy, and Status-Quo Legitimization. The key finding is that a consistent 'lab signal' accounts for significant behavioral clustering — provider-level biases are stable across model updates. This persistence suggests these signatures are embedded in training infrastructure (data curation, RLHF preferences, evaluation design) rather than being model-specific features. The implication is that current benchmarking approaches systematically miss these stable, durable behavioral signatures because they focus on model-level performance rather than provider-level patterns. This creates a structural blind spot in AI evaluation methodology where biases that survive model updates go undetected.