Vigil
Governance and audit for AI in production.
In the constellation
Vigil highlighted in the live map — hover or tap any node to explore.
The gap
Regulated industries are putting AI into decisions that have to be defended — and they can barely govern it once it’s live. They can’t prove why a model decided something, can’t show it would decide the same way again, and can’t see fitness degrading until it’s already an incident.
Auditors ask for a reproducible decision trail. Most teams deploying AI simply don’t have one.
What Vigil is
Vigil is governance and audit for AI in production. It watches deployed models for drift, tracks their fitness over time, and captures a reproducible trail for every decision — so any past decision can be re-run and proven to reproduce. It exports an audit an outside reviewer can actually trust.
It turns “show me why the AI decided this, and prove it would decide it again” into a question with a continuous answer — for the whole deployed system, in a form regulators accept.
What it does
Drift detection
Watches deployed model behavior and flags when it starts to wander from its established baseline.
Reproducible decision trails
Every model decision can be replayed and shown to reproduce — the evidence auditors ask for and rarely get.
Fitness over time
Tracks model quality continuously so degradation is visible, dated, and caught before it becomes an incident.
Auditor-ready exports
Produces compliance records with provenance on every entry — built to be handed to an external reviewer.
The line between judgment and machinery
AccelMars draws one hard line through every product: what an AI decides, and what runs deterministically. Vigil sits on the boundary — and keeps the two honest.
What the AI judges
- Whether observed behavior has meaningfully drifted
- How a decision maps to policy and intent
- Which signals warrant a human review
What stays deterministic
- Capturing the full, reproducible decision trail
- Drift metrics and fitness tracking
- Replaying any past decision on demand
Compliance rules are evaluated deterministically and the decision record is reproducible by construction — the interpretation that calls something a violation stays with policy and a human.
Who it’s for
- Financial services, insurance, and healthcare deploying AI into defensible decisions
- Compliance and risk teams accountable for AI behavior
- Anyone who has to justify an AI decision to a regulator or auditor
Planned — designed, not yet built. It builds on the model-evaluation and drift engines, the reproducible-replay layer, and a new policy engine.