Model Registry and Verification

Model governance in Faramesh is about execution trust, not benchmark marketing. You register exactly which model identity is allowed for a workload, then continuously verify runtime claims against that registry. If an agent silently drifts from an approved model to a different snapshot, or a provider alias changes under the hood, Faramesh gives you evidence and a decision surface instead of blind trust.

In practice, model drift creates two classes of incidents: behavioral drift (same prompts, different outcomes) and compliance drift (a now-unapproved model handles regulated data). The registry helps you pin expected identities per workflow, while verification gives you ongoing proof that the model in use is still the one you approved during risk review.

A good operating pattern is: register approved model identities during change management, verify continuously in runtime checks, and wire failures into incident playbooks. Treat model identity mismatches exactly like identity or signature mismatches in other critical systems.

When to enforce model checks

Use strict verification when model behavior can trigger sensitive side effects or regulated data handling.

Financial actionsRefunds, credits, and approvals require deterministic governance and explicit model approval.

PII or PHI processingData handling policies often depend on approved providers, regions, and model versions.

Production autonomyAutonomous agents should fail closed when runtime model identity diverges from policy assumptions.

Post-incident recoveryModel verification confirms rollback to known-good model baselines after containment.

Register model

bash

faramesh model register gpt-4o --provider openai --version 2026-03 --fingerprint sha256:...

Verify model

Check that runtime model identity matches what was registered. Verification failures should be treated as policy drift, not as minor warnings.

bash

faramesh model verify --agent support-bot

Consistency check

Compare model identity across environments or workloads to catch staging-production skew, unreviewed model upgrades, and hidden provider alias changes.

bash

faramesh model consistency --agent support-bot --window 24h

List models

Show all registered model identities and status.

bash

faramesh model list

Operational pitfalls

Most model governance failures are process failures. Avoid these common mistakes:

!Using provider aliases (for example, generic latest tags) as if they were immutable.

!Registering model identity once and never re-verifying in long-running services.

!Treating verify mismatches as informational logs instead of enforcement events.

!Approving a model for one workload and reusing it across higher-risk workflows without review.

Model verification closes a hidden control gap: it proves the runtime model identity that produced governed decisions, not just the policy text that should have applied.

Need help? Use any of these support options.

open an issue join slack community/forum

MIT License

docs github community get started →