Empirical evidence & governance archive

Case studies & the governance archive.

A public, reproducible record documenting practical approaches to runtime policy enforcement, auditability, decision provenance, and institutional trust for advanced AI systems.

Records
5
Method
Forensic replay
Access
Public
§ 01Incident archive

Cases, modelled against the engine.

Each record pairs a documented historical failure with the deterministic control that would have contained it.

Reproducible
C-0012026-06-08

Case Study 001: Authority Overreach Prevention

Summary

In 2012, a trading firm lost $440 million in 45 minutes due to unintended reactivation of deprecated code.

How Anchor Responds

AST scanning detects unauthorized execution path reactivation; the runtime sandbox (Diamond Cage) isolates and blocks anomalous behavior; the Decision Audit Chain logs approval history.

Modeled | Lessons Applied in Anchor v5.x
Runtime EnforcementIncident AnalysisPrivilege IsolationFinancial Systems
Read case study
C-0022026-06-08

Case Study 002: Policy Drift Management

Summary

In 2024, Air Canada was held legally liable after its customer support chatbot drifted from official bereavement fare policy and fabricated a retroactive refund procedure (Moffatt v. Air Canada, 2024 BCCRT 149).

How Anchor Responds

Anchor's semantic interceptor validates LLM output assertions against the active policy contract before client delivery. Drifted claims are coerced to pre-approved compliant responses and logged in the Decision Audit Chain.

Completed Reference Case | Anchor v5.x
Policy DriftModel ValidationRuntime ContainmentVersion Control
Read case study
C-0032026-06-08

Case Study 003: Audit Reconstruction

Summary

Malicious or security-compromised embeddings introduced into a vector store lead to incorrect but policy-compliant agent decisions.

Combined Solution

Canon ensures source material freshness and authenticity, while Anchor enforces runtime validation of LLM output against the current active policy manifest.

Active Research | Canon + Anchor
Forensic AuditingCryptographic LedgersDecision TraceTelemetry
Read case study
C-0042026-06-12

Case Study 004: TSB Bank IT Migration (2018)

Summary

In 2018, TSB Bank's IT migration failure caused a £600M outage affecting 1.9 million customers, resulting in a joint FCA/PRA fine of £48.65M.

How Anchor Responds

CI/CD defect limit gates (POL-DEPLOY-004) enforce active-active datacenter configuration drift detection before production cutover. The runtime sandbox blocks deployment when unresolved defect counts exceed constitutional thresholds.

Completed Reference Case | Anchor v5.x
Configuration DriftCI/CD EnforcementIntegrity VerificationOperational Resilience
Read case study
C-0052026-06-12

Case Study 005: Citibank Revlon Transfer (2020)

Summary

In 2020, Citibank accidentally wired $893M to Revlon lenders due to an Oracle Flexcube UI checkbox failure, resulting in a $500M net loss and OCC Consent Order.

How Anchor Responds

Structured Mode payment rules (POL-FIN-005) enforce multi-party approval gates and amount-threshold verification before high-value wire execution. The Decision Audit Chain captures the full authorization lineage.

Completed Reference Case | Anchor v5.x
Structured ModeAuthorization ControlsNo-Prose RuleTransaction Verification
Read case study
§ 02Multi-framework evaluation pipeline

Upcoming assessments & active audits.

We regularly run Anchor's frozen engine checks against mainstream autonomous frameworks to audit governance boundaries, capability leakage, and unsandboxed execution limits.

CompletedJune 2026

OpenHands Runtime Governance Audit

Analysis of process-level Docker boundaries, AST parsing triggers, and the semantic coupling of core message contracts in production integrations.

Report pending final review
Under reviewJune 2026

crewAI Boundary & SDK Resolution Audit

Static capability check evaluating direct public API initializations, execution delegation tradeoffs, and local CLI subprocess sandboxing.

Review in progress
QueuedQueue active

LangGraph Cyclic State Checkpointing Audit

Auditing of state transitions, local execution parameters, and simulation environment governance boundaries in graph configurations.

Scheduled