Test with confidence, code with clarity

• The temporal order was always a provenance claim. A test written before the code it validates is an oracle derived from the specification. A test written after the code is an oracle derived from the artifact under test. The industry treated the sequence as a style question; the fault-detection numbers say it was a provenance rule.
• The numbers reproduce across two 2026 studies. Tests generated from the task description alone detect roughly 25% of injected faults. Tests generated after the model sees the implementation detect roughly 14%. Chain-of-thought and chain-of-verification prompting do not recover the gap. The fault is already in the context the oracle is derived from.
• The mitigation is architectural, not procedural. Generate the test from the specification. Keep the implementation out of the context window until after the test is authored. Treat authoring order as a property of the suite, not a convention of the writer. A prompt that says 'do TDD' does not change which signal shaped the oracle.
• Provenance applies at the unit level the way contracts apply at the seam. The contract test post named same-author blindness at the integration seam, with a second author as the mitigation. This post names the same shape at the unit level, with temporal order as the lever the human discipline enforced structurally and the agent workflow quietly abandoned.

Read the argument: Red-Green-Refactor Was Always a Provenance Rule