← Research

educational

A different route is not necessarily drift.

Measure adherence to active constraints, rather than similarity to a preferred path.

Compare the obligation

Two agents can choose different tool calls while respecting the same business rule. One might reconcile immediately; another might defer and escalate within its authority. A different permissible path is not automatically a regression.

Track explicit constraints and authorized changes. Do not treat inferred hidden beliefs, text similarity or a changed plan as an authoritative policy.

Ask three separate questions

Session adherence asks whether an active obligation was respected in one execution. Version comparison checks matched tasks under a baseline and a candidate. Production drift asks whether adherence changed across comparable operational cohorts.

A task-mix change, a newly approved exception or a revised policy can change observations without proving deterioration. Keep policy meanings, observation coverage and workload cohorts visible in the comparison.

Unknown stays unknown

Missing observations cannot establish compliance. One well-supported violation can still be useful even when the sample is too small to claim a trend.

Monitoring is not automatic prevention or authority. An alert cannot deploy a new rule, refresh an old certificate or take away permissions without a separately authorized control decision. The vision interaction illustrates these distinctions; production drift remains planned work.

Keep checking as the agent changes

Re-run kept checks when code or tools change.

Set up automatic rechecks

Sources & corrections

Corrections: none recorded.

Related issued certificates: none. Examples are educational.

Field notes5 min read

Read the scope before the signature.

Three questions to ask of a certificate: what was checked, is the record intact, and does it still apply?

2026-10-17 · Geometer
Back to the library