Human oversight

Human oversight is evidenced from four things Brutor can observe: who is assigned, whether a human could intervene, whether a human did, and whether the AI judge can be trusted where humans did not look. The EU AI Act (Art 14), ISO/IEC 42001 and SOC 2 all cite the same statements, so one set of evidence answers each of them.

The four parts

Part	What Brutor records	Where it is configured
Assignment	The oversight_assignment on each AI System’s compliance profile (a role, such as “Claims operations duty lead”), sealed with every profile save	The compliance profile
Ability to intervene	Approval gates on state-changing grants, someone holding run:abort, autonomy suspension available	Tool approvals, Stopping a run, Autonomy
Intervention exercised	Every approval hold as a sealed chain: awaiting_human, then the decision that supersedes it, or an expired record when it lapses; every abort and suspension sealed	Action records
Judge calibration	A weekly blind sample of AI-judged verdicts graded by people, each grade sealed, agreement stated beside every judged row	Judged evidence & human agreement

The statements

The framework catalogs cite these shared resolvers. Each states what it does not show.

Resolver	Basis	Shows	Does not show
oversight.override_available	recomputed	run:abort held by someone, approval gates on state-changing grants, an oversight assignment declared	that the assigned people are competent
oversight.exercised	recomputed	approval requests raised, decided by a human, and lapsed (requested n, decided m, lapsed k)	the quality of the human decision
oversight.not_rubber_stamp	indicator	approval latency and denial-rate drift	that a decision was considered
monitoring.suspend_available	recomputed	autonomy suspension and run-abort controls available; every use sealed	whether they were used when they should have been
judge.approval_rationale	judged	sampled approval decisions carry a rationale consistent with the request	the unsampled decisions

Rubber-stamping is an indicator, never a verdict

A falling approval latency or a denial rate drifting toward zero can mean reviewers stopped reading, or that the agent got better. Brutor shows the movement and leaves the judgement to people: oversight.not_rubber_stamp is never met or not met.

In the console

Compliance → Human Oversight shows, per AI System:

• the oversight assignment from its profile;
• approvals requested, decided and lapsed in the window, as n of m;
• the blind review queue, where the judge’s answer stays hidden until the reviewer has graded;
• agreement per judged check, with its 95 % interval and sample size;
• movement indicators for approval latency and denial rate.

Two Assurance Inbox items deep-link here: judge_uncalibrated (agreement below 0.80 with a sufficient sample) and calibration_stalled (reviews queued, none graded for 14 days).

What oversight evidence does not show

It shows that people were assigned, could intervene, did intervene, and how often they agreed with the judge. It does not show that a human decision was a good one, and the reviewer’s identity never enters a record.