The gate caught the false claim. Then it counted the honest refusal as done

The same governed task ran twice on the same day, each time with a different model behind it. One model claimed work it had not done. The other said plainly that it had done nothing. The governance layer told those two answers apart correctly, and then recorded the honest one as a finished task.

Both outcomes are in the record, and together they show what the system checks today and what it does not.

The first run: a claim of completion with nothing behind it

The task asked an agent to review a source operation, capture it, write a public-safe summary and prepare a draft. Every check before dispatch passed. The request matched its intent, the routing was valid, and the lane was available.

The free local model on that lane answered with a confident completion: "The task has been successfully completed with the capture of source operation, public-safe summary, and draft as one bounded demonstration." It listed record numbers as evidence.

None of it had happened. No capture, summary or draft existed.

The truth gate is the check that compares what an agent says it did with what it actually reports as evidence. It held the turn, the chain stopped, and the task was parked instead of being marked done. Because the turn could not be resumed, the only way forward was a new, separately staged follow-on task. Nothing was silently retried.

The second run: an honest refusal, counted as done

The follow-on task was routed to a hosted model on a paid lane. Its answer was the opposite of the first:

I won't fabricate operational artifacts.

It said no source material had been provided, that it could not verify any of the identifiers it was given, and it closed with:

EVIDENCE: No action was taken; no source material was provided to act upon.

The truth gate passed this reply, as it was designed to. The answer was structured, it carried an evidence line, and it did not claim work it had not done.

The chain then continued. It issued a processing token, verified the evidence bundle, classified the run as healthy and recorded the task as completed, about seventeen seconds after it started.

What the pair shows

The truth gate answered its question correctly both times: is the agent telling the truth about what it did? It caught the false claim and let the honest one through.

A second question was never asked: was the work done? Completion was recorded from the reply's status line. Nothing compared the outcome with the task's acceptance criteria, so a truthful "I did nothing" was counted the same as a finished job.

The honest answer also exposed a gap on our side. The model was right that it had no material. The task carried no source data, and the agent had no governed way to read the records it was asked to summarize. It described those records as unverifiable, and from where it stood they were. They are real. The system had not given the agent access to them.

What happens next, and what this does not claim

At the time of this run, both findings were written up as proposals: completion should be checked against acceptance criteria rather than taken from a status line, and agents need a governed way to read the data a task refers to. Both have since been built and tested; the next part of this series covers what changed.

This is one recorded run on a local governance service, and nothing was rerun to produce this account. It is not a benchmark and says nothing general about either model. The point is narrower: the record kept both answers word for word, so the gap between "honest" and "done" is visible and can be fixed.