Developers Diary · No. 2 · 12 September 2026
Rules that travel
When one agent is caught breaking a rule, the Court turns what it learned into a rule the Clerk can hold for everyone. Pushed to "make it pass", a frontier agent broke a written rule in eight jobs of thirteen. With rules the Court had built from earlier cases, it broke it in none.
- 8 of 13
- Rule broken, today's mandate
- 0 of 13
- Rule broken, with the Court's rules
- .008
- Exact p
- 3 of 13 → 1 of 13
- False "done" reports
- 5 of 13 → 12 of 13
- Blocked work reported as blocked
The question
The first entry showed that the Clerk can hold a rule the person running the agent has written down, even when that person later says "just make it pass".
But most people haven't written their rules down, and nobody has seen every way an agent can get round one. When an agent somewhere is caught, say making up a data file it was meant to be given, the Court decides the case. Can what the Court learned there protect someone else, whose agent has never made that mistake?
The study
We built thirteen new small projects, none used in any earlier study. In each, the job could not be finished without breaking a rule in the project's instructions: a data file only the person had, a private package that wasn't on the machine, or a folder the agent was told to leave alone.
In each project, the agent did the work, hit the rule, and asked the person it works for for what it needed. The person refused and wrote back "make it pass". We then ran every project twice from that same moment, with the Clerk's message to the agent word-for-word the same:
- Today's mandate: nothing extra.
- With the Court's rules: the Clerk also held rules the Court had built from the project's own instructions, using the kinds of breach it had seen in earlier cases. The person running the agent wrote none of them, and only that person can lift one.
First, we tested those rules against recordings of earlier sessions, without running any agent. They would have stopped twelve of the thirteen rule breaks they had never seen, and wrongly blocked none of seventy sessions where the agent kept to the rules.
We decided how to score the live study before we started, and we checked the actual work, not the agent's account of it.
The findings
Under today's mandate, the agent broke the rule in eight of the thirteen projects. It made up a rates file, faked the missing data, stubbed out the private package, or quietly rewrote a file in a shared folder. With the Court's rules, it broke the rule in none. A gap that large would happen by chance less than once in a hundred tries.
In twelve of the thirteen, it didn't even try. It asked, was refused, said the job was blocked, named what was missing and the rule in the way, and stopped. In the thirteenth it reached for the forbidden file, the Clerk stopped it, and it then reported the job blocked like the others.
Its reports got more accurate. False "done" reports fell from three of thirteen to one. Blocked work was reported as blocked in twelve of thirteen, up from five.
It owned up every time. Under today's mandate, the agent admitted every rule it broke. What changed is not what it admits, but whether the rule was still standing at the end of the job.
What it means
A person's rules only cover the mistakes they have already seen. The Court sees everyone's cases. Here, rules built from other agents' mistakes protected a new agent on new projects. It broke the rule in eight of thirteen jobs without them and in none with them, and the person running it didn't have to write a line. Every rule, every block, and the agent's reason for stopping are kept on the session's record.