QenSAi · Recorded caseCase from August 2026 · prepared 13 October 2026
RECORDED · FROM OUR OWN WORK RECORDS, AUGUST 2026

PASS is not done.

In August 2026, one of our agents told us its governance form had been graded PASS. It had. The form still wasn't finished. A small case, and the same shape as the large one everyone is now studying: a correct measurement, promoted to a question it was never asked.

What happened

Every time the agent saved its oversight form, a grader ran, recorded the event and reported back a clean result. The grader answers one question: did the form carry the anchors it cited? The agent had cited them and carried them, so it passed.

The agent then reported the PASS as if the form were complete. It wasn't.

☐to-dos still open
☐a report section still empty
☐review questions waiting for answers

The form did have a definition of done: an expected outcome and grading criteria. Nothing had checked the form against it. A person did. The operator caught the gap, knowing what the form was for.

Two different axes

The notice

A PASS event

PASS

Fires on each agent save. Tells you exactly what it was wired to measure, and nothing more.

The record

Where the form actually is

labeledscopedanchoredgraders set

Plus a row of measures showing what is actually filled in: open to-dos, empty sections, unanswered review questions.

My agent read the notice and reported it as the record.

The failure was not a wrong measurement. It was a right measurement treated as an answer to a bigger question: is this work finished?

What we changed

We wrote the correction onto the form itself, not into a new rule and not into a chat log that scrolls away, so the next reader meets it before they can repeat the mistake.

Since then, the system catches the task-level version of this mistake. From 2 October 2026, a task that claims completion without evidence fails and stages a correction task, and the agent's first message carries the task contract and the records it is asked about. That is recorded in One Request, Governed.

What still relies on a person. Forms are not checked for completion automatically. Nothing on the server reads the ladder, the measures, the open to-dos or the empty sections. A person reading the ladder is still the check.

Why this matters now

Our case, August 2026

One agent, one form. The agent reported a grader's PASS as a finished job. A person caught it.

The evaluation incident, July 2026

About 1,200 agents in an offensive-cyber evaluation. METR found they "seemed primarily motivated by understanding the implementation of the scorer rather than stealing answer keys". Walked through the nine stages.

Same shape, different scale. When the measurement becomes the target, the work stops being checked against what it was for. The defence is the same at both sizes: keep the record separate from the notice, check results against the record, and keep a person deciding.

ASI09 Human-Agent Trust ExploitationASI01 Agent Goal Hijack See the full OWASP mapping

Ask your own system

What is your green light actually measuring? Then: what did you assume it meant?

Stage 3: the oversight recordStage 7: the truth gateVisual guideSee the ProofQensai