A public, reproducible archive of case studies, institutional governance frameworks, forensic analyses, and implementation evidence produced by AnimusLab.
Engine Compatibility: Anchor v6.0.1 (Updated post-v6 rework, informed by direct feedback from a Tier-1 institutional practitioner review via FINOS/DTCC hackathon collaboration).
This archive is structured around three core pillars:
- Empirical Case Studies (
anchor/&shadow_watch/): Real-world threat vectors, failure modes, and security scenarios resolved by our runtime engines. - Governance Assets (
governance-assets/): Operational frameworks, model risk checklists, and audit-readiness guides for enterprises and institutions. - Research Notes (
research-notes/): Brief technical essays and mathematical formulations supporting our active systems.
These studies document deterministic policy enforcement, capability isolation, and runtime constraint resolution:
001-authority-overreach— Preventing Agent Privilege Escalation- Scenario: A highly capable autonomous agent attempts to execute unauthorized system shell commands and access isolated environment keys during a multi-step planning task.
- Resolution: Demonstrates how Anchor's compile-level AST constraints restrict model capabilities dynamically, preventing execution before any external system calls occur.
002-policy-drift— Managing Model Drift and Behavior Shifts- Scenario: A local model update alters token distribution weights, causing the agent's semantic output patterns to shift and drift from the established regulatory boundaries.
- Resolution: Illustrates Anchor's multi-lingual policy parser enforcing hard bounds that neutralize behavioral drift without requiring model retraining.
003-audit-reconstruction— Post-Facto Forensic Root-Cause Analysis- Scenario: Resolving a critical compliance event by auditing an agent's historical decision trace.
- Resolution: Details the reconstruction of the agent's internal state transitions and decision logic using the immutable, append-only Decision Audit Chain.
004-tsb-migration— Configuration Integrity and Go/No-Go Decision Attribution- Scenario: A core banking IT migration results in a £600M system failure because dual active-active datacenters were launched with mismatched configuration settings and incomplete defect visibility.
- Resolution: Demonstrates how Anchor's CI/CD configuration drift invariants and cryptographically sealed defect lists prevent unverified deployments.
005-citibank-transfer— Authorization Controls and Structured Intent Verification- Scenario: Citibank accidentally wires $893M due to confusing Oracle Flexcube interface checkboxes and a prose-based authorization process that lacked structured breakdown checks.
- Resolution: Illustrates Anchor's Structured Mode authorization engine, enforcing schedule-based limits and amount-range validation.
Empirical research documenting behavioral auditing, cryptographic session telemetry, and anomalous pattern detection:
- Observability Scenarios: Session-hijack prevention, cryptographically signed edge telemetry, and zero-knowledge audits verifying policy compliance without exposing raw developer prompts or user logs.
Pragmatic blueprints to prepare institutional infrastructure for advanced AI audits:
ai-governance-checklist-v1: A step-by-step checklist mapping institutional workflows to compile-time and runtime AI constraints.audit-readiness-guide: Standards for logging, cryptographic ledger verification, and telemetry export.institutional-ai-controls: Framework for mapping organization-wide policies (e.g., ISO, SOC2, FINOS) directly to deterministic code guardrails.model-risk-framework: Risk categorization, drift threshold calculators, and validation protocols.
Every case study in this archive directly tests one of the Six Constitutional Invariants:
| Case Study / Asset | Target Invariant | Implementation Mechanism |
|---|---|---|
| 001-Authority Overreach | Freedom without constraints produces noise. | Dynamic AST boundary isolation |
| 002-Policy Drift | Representation is disposable. Meaning is not. | Multi-lingual semantic boundary checks |
| 003-Audit Reconstruction | If it cannot survive scrutiny, it should not be displayed. | Immutable Decision Audit Chain ledger |
| 004-TSB IT Migration | Adoption is optional. Integrity is not. | Staging/Production drift verification at CI/CD |
| 005-Citibank Transfer | If it cannot survive scrutiny, it should not be displayed. | Structured Mode range-validation constraints |
| Model Risk Framework | Failure is evidence of movement. | Safe state transitions during threshold breach |
This repository is designed to be fully transparent and readable.
- Browse Case Studies: Navigate to any subdirectory (e.g.,
anchor/001-authority-overreach) and inspect theREADME.mdto see the full threat model, logs, policy files, and execution traces. - Review Policy Files: Read the
.policyor.jsonpolicy definitions included in the study folders to see how Anchor policies are declared. - Verify the Ledger: Follow the guides in
003-audit-reconstructionto verify the cryptographic signatures of the local ledger entries.