Case Studies &
Governance Archive
A public, reproducible record documenting practical approaches to runtime policy enforcement, auditability, decision provenance, and institutional trust for advanced AI systems.
Case Study 001: Authority Overreach Prevention
In 2012, a trading firm lost $440 million in 45 minutes due to unintended reactivation of deprecated code.
AST scanning detects unauthorized execution path reactivation; the runtime sandbox (Diamond Cage) isolates and blocks anomalous behavior; the Decision Audit Chain logs approval history.
Case Study 002: Policy Drift Management
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).
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.
Case Study 003: Audit Reconstruction
Malicious or security-compromised embeddings introduced into a vector store lead to incorrect but policy-compliant agent decisions.
Canon ensures source material freshness and authenticity, while Anchor enforces runtime validation of LLM output against the current active policy manifest.
Case Study 004: TSB Bank IT Migration (2018)
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.
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.
Case Study 005: Citibank Revlon Transfer (2020)
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.
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.
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.
OpenHands Runtime Governance Audit
Analysis of process-level Docker boundaries, AST parsing triggers, and the semantic coupling of core message contracts in production integrations.
crewAI Boundary & SDK Resolution Audit
Static capability check evaluating direct public API initializations, execution delegation tradeoffs, and local CLI subprocess sandboxing.
LangGraph Cyclic State Checkpointing Audit
Auditing of state transitions, local execution parameters, and simulation environment governance boundaries in graph configurations.