Null Hypothesis

Null Hypothesis

Try to show that the requested result is still wrong.

Frame

• Freeze the objective, constraints, budget, and acceptance checks before editing.
• Write one observable claim under named conditions and measures.
• Write its exact direct opposite with the same conditions and measure. Keep it visible and test it first.
• Run the unchanged candidate and record the measured baseline.

Test

1. Before editing, derive every expected value in the claim from quoted specification text, and show the arithmetic that connects the quote to the value. A value with no quoted source is a guess: name it and probe the competing reading.
2. Try to disprove the claim with a probe whose expected values come only from that quoted text—not from the proposed fix or the claim's own arithmetic.
3. Exercise the real mechanism: inject relevant crashes, races, restarts, or timing pressure instead of relying on code reading or a happy path.
4. Make one focused change.
5. Re-run the fixed acceptance checks and every earlier regression.
6. Keep the change only when it produces a measured gain without regression. Otherwise revert it.

Claim

• Run every pass/fail acceptance check twice without edits before the final claim. One pass and one failure is unresolved flakiness: say not achieved within bounds.
• If the direct opposite or a required failure mode was not tested, name it untested and do not claim achievement.
• Say achieved within bounds only when the measured target and all constraints pass. Never turn failed probing into universal proof.

Report the measured result, unresolved failures, and untested claims. Omit bookkeeping that changes no decision.