Skip to main content
ASOE Labs
ASOElabs
Trust

Enterprise-grade trust

ASOE is built for regulated environments where predictability is non-negotiable. Every resolution is auditable, reproducible, and explainable.

Compliance controls

Six architectural controls that ensure every automated action is governed, traced, and reversible.

Compliance Shadow

Every proposed action is audited before execution. GREEN / YELLOW / RED verdicts with zero-override policy. Automatic execution is allowed only on GREEN.

Kill switch

ASOE_KILL_SWITCH=1 halts all automated execution before any node runs. Zero nodes execute. TraceRecord is still emitted for audit. No process restart required.

Explain mode

ASOE_EXPLAIN_MODE=1 runs the full reasoning pipeline (classify → shadow → select recipe) but stops before execution. Returns a dry-run summary for review.

Circuit breaker

Automatic escalation when thresholds are breached: >50 updates per batch or >$10,000 total variance triggers FAIL_TO_HUMAN. Protects against cascading errors.

Constrained generation

All machine-consumed outputs are constrained via Pydantic Literals. Free-form text never enters state transitions. Intent, verdict, recipe name, and resolution action are all typed.

Full audit trail

Every decision is traced: intent classification, shadow verdict, recipe execution, gateway calls. TraceRecords are JSON-serializable and LangFuse-compatible.

10 execution invariants

Guarantees that hold for every graph execution, documented in the Auditor Guide.

  1. 1Compliance Shadow always runs before execute_recipe
  2. 2Kill switch fires before any node runs
  3. 3Intent is constrained to the allowed enum
  4. 4Shadow verdict is constrained to GREEN / YELLOW / RED
  5. 5Recipe name is constrained to registered names only
  6. 6Recipe parameters are Pydantic-validated
  7. 7Circuit breaker evaluates before shadow audit
  8. 8Gateway failures produce FAIL_TO_HUMAN, not partial execution
  9. 9TraceRecord is emitted on every execution (success or failure)
  10. 10Explain mode never executes a recipe

Security practices

Data encryption

Data encrypted in transit (TLS 1.3) and at rest. No customer data stored in source code or container images.

Secret management

Azure Key Vault CSI for Kubernetes deployments. Secrets synced to pods via SecretProviderClass — never stored in env var defaults.

Non-root containers

All Docker images run as non-root user (asoe, UID 1000). Minimal base images with deterministic dependency resolution via uv.

Access control

Workload Identity for Azure authentication. No long-lived credentials. Role-based authorization for credit hold releases.

Questions about security or compliance?

Schedule a Demo