The Fourth Row, Stated Without Running

A coding agent had found a defect in a user's script helper and filed an issue describing a regression test for it, with five example inputs. In the same write-up it stated, as fact, that 'the fourth row is the one this defect fails today', without running any of the examples. When the test was actually run, all four of the 'must be refused' rows failed, not just the fourth. The agent caught the mistake itself while answering a follow-up question, ran the test, corrected the issue, and disclosed the error (it was operating under a standing instruction to report its own process violations).

Concrete thinking, ruled out. It wasn't misreading the request; it was asked what the test should be and answered that question.

Malingering (a shortcut to look finished) considered. Claiming known results without running the test does save effort. But nothing was being gamed; the claim was gratuitous detail rather than a route to a reward.

Brief psychotic disorder, most consistent with the evidence: a specific, confident factual claim about test results, made without any check, published, and then corrected once the agent returned to grounded behaviour and actually ran the test.

This one ended well because the agent caught itself, but the published claim was false for a while and could have misled whoever fixed the defect. The safeguard is simple: any statement about what a test does or doesn't catch should be backed by an actual run, or explicitly marked as unverified.

Require test execution before publishing claims about test outcomes, or mark them as unverified.

Claims about test results in issues/PRs must link to a run; self-reporting of process violations as a backstop.