The Control Platform

Glacis’s earlier product materials describe policy evaluation and attestation at governed AI boundaries, including proxy-mediated model requests. Use that architectural description within its documented scope. A deployment may also have tool, application, approval and downstream effect boundaries. GROUNDED · Labs

Three jobs

Job	Scoped account	What still needs verification
Measure	Record declared events, context and comparisons	Eligible population, completeness, detector validity
Enforce	Apply a named rule on a mediated path	Installed configuration, bypass resistance, effect boundary
Attest / witness	Sign a bounded record and optionally witness its commitment	Trust model, keys, retention, verification and missing records

These are capability categories, not an assertion that every deployment implements them identically. STABLE

The boundary is the claim

A model proxy can observe calls routed through it. It cannot infer every tool action, user decision or downstream effect from the model request alone. An agent tool running outside that path needs separate mediation and observation if the claim concerns its action. STABLE

A deterministic policy evaluator can still receive wrong inputs or an inappropriate rule, and can be bypassed if alternate routes exist. A base-URL change can suit a compatible proxy integration; it is not a universal integration guarantee. STABLE

Describe capabilities as proposed, implemented, tested, deployed and accepted for a defined use. Claim prevention only when the relevant action or output was stopped before its effect in the tested/deployed scope. See controls and blind spots.

Next

→ Baselines and drift

Drill this page →21 bank questions stand behind what you just read. Check it while it’s warm.