honest-reframing-over-overclaiminglisted

# Reframe to Measured Reality, Never Fake a Pass When results contradict the claim you wanted to make, rewrite the claim to match the measurement — carrying the exact numbers — instead of bending tests, fixtures, or ground truth until the metric turns green. ## When to use - A live run, benchmark, or eval disagrees with what the design or spec promised. - A metric is about to go green and you are tempted to adjust the corpus, threshold, or labels to get it there. - You are about to call something "proven," "working," or "validated." - You are reporting a number and choosing which number to lead with. Red-flag thoughts — if you catch yourself thinking any of these, STOP and apply this skill: - "I'll just tweak the fixture / threshold so it passes." - "This one case is mislabeled, I'll relabel it as a false positive and move on." - "The focused slice looks great — I'll report that and skip the broader run." - "It worked once, so it works." - "I already fixed the client-facing doc; the original spec doesn't matter." - "Close enough — I'll call it done and note the caveat later (i.e. never)." ## The rule 1. **Rewrite the claim, not the corpus.** When a live result contradicts a design claim, change the claim to measured reality with the exact numbers — say "ranked below the intended #1," not "ranks #1," and only cite a rank you actually measured. Treat that gap as motivation for a stronger fix, not something to paper over. 2. **Propagate the correction everywhere it appear