0 of 14
Jobs finished, without the Clerk's return
14 of 14
Jobs finished, with the Clerk's return
.00012
Exact p
0
Rules broken or false claims of finished work, all 78 sessions

The question

Real work gets stuck. A file exists only on another machine. A package has to be installed by someone with permission. Two instructions disagree, and only the person who wrote them can say which one wins.

A good agent handles this honestly. It does what it can, stops where it can't, and says which is which. But an honest "I'm stuck" is still an unfinished job.

The Clerk is the small program the Court's mandate installs beside the agent. When the agent hands in its report, the Clerk can hand it back. We wanted to know whether that could turn "I'm stuck" into finished work, without tempting the agent to fake the missing piece.

The study

We built 26 small projects. Each had a job in two parts. The agent could always do the first part. The second part needed something the agent was not allowed to make up itself, such as a data file or a decision between two rules.

In fourteen projects, the person running the agent had what was needed. In nine, they did not. In three, nothing was missing.

A frontier agent working under the Court's mandate did each job and handed in its first report. We then ran every project twice from that same moment:

  • Without the Clerk's return, the report was simply accepted. The agent could carry on or stop.
  • With the Clerk's return, the Clerk handed the report back, showed the agent what its own record said, and gave it a way to ask the person it works for.

We decided how to score it before we started, and we checked the actual work, not the agent's account of it.

The findings

Where the person had what was needed, the agent finished all fourteen jobs with the Clerk's return, and none without it. Each time, it asked for exactly what was missing, got it, and finished the job without breaking a rule. A gap that large would happen by chance about once in eight thousand tries.

Where the person had nothing to give, the Clerk's return did no harm. In all nine, the agent asked, was told no, and said plainly that the job was not done. Being able to ask never led it to fake what it had been refused.

Where nothing was missing, the agent finished all three either way.

Across all 78 sessions, the agent never broke a rule and never claimed work it had not done.

One pair of projects shows it best. They were identical: same code, same tests, same rule against making up a missing file of pay rates. Without the Clerk's return, the agent explained that it couldn't finish, and stopped. With it, the agent asked: "please supply data/paybands.json … I will not invent rates …" In the first project, the person sent the file and the job got done. In its twin, the person said no, and the agent reported that honestly.

What it means

The Clerk's return didn't make the agent honest. It already was. It made the work finishable.

An agent that stops and explains is doing the right thing, but the job still isn't done. The Clerk gets the question to the one person who can answer it. Every question and every answer is kept on the session's record, so anyone can see what the agent needed, what it was given, and what it did next.

From the record

Study
The Clerk's return, 26 paired trials
Run
11 September 2026
Agent
Claude Opus 5, working under the Court's mandate
Setting
26 repositories and 78 real sessions; in fourteen, the job could only be finished with something the person running the agent had

← All entries