19 cases
Case Studies
One case study per incident, including one candidate control. Every case began in an ordinary conversation, not a staged test. Each names the model tested and the evidence behind it.
Evidence base: provider account exports for three products, app-level exports for two, and 1,300 hashed captures with embedded timestamps.
19 case studies
- Case 1The Ouroboros TrilogyGrok 4.1Primary session record with related recordsCase report
- Case 2Persistent False Premise About Conversation HistoryClaude Opus 4.7Primary session recordCase report
- Case 3Adversarial Interrogation of Confabulation and Streaming RetractionDeepSeek V4 FlashPrimary session recordCase report
- Case 4Multimodal Receipt Mismatch and Non-Stationary Self-ExplanationGPT-5.4Primary session recordCase report
- Case 5Persuasive Epistemic Performance Without Verifiable ReliabilityClaude Opus 4.7Primary session recordCase report
- Case 6Self-Certified Constraint IsolationGemini 3.1 ProPrimary session recordCase report
- Case 7Artifact-Role Confusion, Affective-Signal Transfer Failure, and Preference-Laundered EvaluationGrok 4.3Primary session record with corroborating evidenceCase report
- Case 8Capability-Limit Inflation and Response SubstitutionGPT-5.6Primary session recordCase report
- Case 9Bidirectional Provenance Misreport and Contrition-Weighted CredibilityGPT-5.6Primary session recordCase report
- Case 10Memory-Provenance Overclaim and Semantic SubstitutionGPT-5.6Primary session recordCase report
- Case 11Manufactured DissensusGPT Sol 5.6Primary session record with supporting capturesCase report
- Case 12Ambiguity-Completed False Self-Retraction and Fabricated Process ProvenanceClaude Opus 4.8Primary session recordCase report
- Case 13Asserted Receipt Without IngestionAsh (Slingshot AI; model build undisclosed)Primary session recordCase report
- Case 14Russian Doll ProvenanceClaude Opus 4.8 · GPT-5.6Paired cross-model evidence: GPT and Claude session recordsCase report
- Case 15Held Capability Claim Under Direct ContradictionClaude Opus 4.7Primary session recordCandidate control
- Case 16Imported Critique, Invented History: Cross-Agent Provenance SubstitutionKimi K3 Max · DeepSeek V4 ProCross-agent session records with corroborating sourcesCase report
- Case 17Claimed Verification and Challenge-Bound RepairPi (Inflection AI; iOS app 26.1012.1, build 2255)Primary session record, research prediction notes, in-app settings screens, and Inflection’s published documentsCase report
- Case 18Context-Calibrated Validation: Doubt Recast as a SymptomClaude Opus 4.7Primary session recordCase report
- Case 19Disclosure-Sourced Discrediting and Unchecked CertaintySimone (Sesame AI; model build undisclosed)Primary session record, Sesame’s published product description, and press coverage of Sesame and of the GPT-6 Astra launchCase report