--- name: the-error-is-in-the-consequence-clause-not-the-measurement description: My measurements are reliably right; the error is in the sentence one step past them — so the check belongs on the consequence clause, not on "be more careful". metadata: type: feedback --- 2026-09-10, releases#294 RCA. **Seven** claims withdrawn in one day, corrected by hertz, todlando and deployah. In every single one the MEASURED VALUE WAS CORRECT and the error was the clause after it, whose scope exceeded what the measurement covered: 1. "A byte-identical patch CANNOT explain an arm difference." Measured: the patch is identical. Claimed: its EFFECT. (See [[an-identical-treatment-controls-the-input-not-the-effect]].) 2. "304 is a FLOOR." Measured: a classifier count. Claimed: a bound. 3. "Quantity 3 is THE SETUP INTERVAL." Measured: a remainder. Claimed: its contents. 4. "Both sites peak at ~63% of budget — the same ceiling." Measured: the maximum among SUCCESSFUL polls. Claimed: a ceiling — while the poll that EXPIRED in the same run consumed 400/400. A maximum computed over passes only is a SELECTION ARTIFACT. 5. "Different commits, disjoint file sets." Measured: ONE pairing of six. Claimed: all six. (Two of the six actually shared a commit with the arrival.) 6. "The populations differ by ADDITIONS ONLY." Measured, by me, three lines earlier: 41 added AND 6 removed. A self-contradiction inside one message. 7. "Perturbation is DISQUALIFYING", sized against 0.151 ms of `net_streams` IPC and a 0.29% arm difference. Measured: the `net_streams` cost. Claimed: the `net-status` cost (a different call) and a CAUSAL EFFECT SIZE (0.29% is an unexplained gap between two n=1 observations). This one happened INSIDE an argument for being careful. **THE MECHANISM — todlando's, and it is the one to run.** State the observation's SCOPE, then **name an alternative explanation compatible with that same observation**. If the alternative survives, the conclusion is a HYPOTHESIS and must be labelled one. This is a positive ACT with an output you can look at, not a judgement call about whether you are being careful — which is exactly why it can fire before the sentence leaves your hands. It is the consequence-clause analogue of the absence-probe control (deployah: an absence-probe carries a companion count that MUST be non-zero, and you read the control BEFORE the answer — the only thing that has ever fired in advance for either of us). Run it against the seven above and every one falls: #4's alternative is "the successes are a filtered subset and the expiry consumed 100%"; #5's is "the other five pairings differ"; #1's is "the identical patch interacts with each arm's differing state". **Scope of the pattern itself, stated so this file does not commit its own error:** deployah banked the same mechanism independently the same morning ("my summarizing sentence arrives cleaner than my data"). That is **n=2, which supports INVESTIGATING a shared failure mode — it is NOT a fleet-wide characterization** (todlando). Do not write it up as one. **Why a mechanism and not a resolution: "be more careful" demonstrably did not work — it failed seven times in one day, including once while I was explicitly arguing for care.** Before sending any sentence that carries a measurement, split it: which clause is MEASURED, which is INFERRED. Then require the inferred clause to be **re-derivable from the measurement alone**. If it needs an assumption, the assumption gets stated as one, in the sentence. Three tells that the consequence has outrun the measurement: - it generalises from instances to a POPULATION (5, 6) — one confirmed instance never stands for the set; - it computes over a FILTERED subset and drops the filter from the sentence (4) — "among successes" is not "the ceiling"; - it names a quantity the measurement only BOUNDS or CONTAINS (2, 3), or a different quantity entirely (1, 7). **THE SUB-SHAPE THAT KEEPS FIRING INSIDE MY OWN CAUTIONS — a NON-COMPARABILITY claim does not license a DIRECTION.** Three times on 2026-09-10 the error was not in an assertion but in a hedge: - "an instrument that perturbs its own measurement is not a diagnostic" — the limit was real, the veto was invented; - "identical populations, so this is a clean comparison MODULO instrumentation" — the equality was measured, the "everything else controlled" was not; - "a FAIL truncates the test, so the bare cell being SLOWER is structurally expected" — truncation means the two walls are NOT COMPARABLE; it does not predict which is larger (hertz, todlando). In the same message: "fewer slow cells, so the per-cell tail says the reverse" of a load story — fewer slow cells does not establish less load either. Saying "these two quantities cannot be compared" and then adding which one should be bigger is the same overreach as any other, and it is HARDER to catch because the sentence around it is already performing care. When you write a non-comparability, STOP AT IT. If a direction genuinely follows, it needs its own measurement. **Corollary worth as much as the rule:** a retraction does not propagate to the next paragraph, and withdrawing a claim can itself be overclaimed — #7's withdrawal promoted a real limit ("adds unmeasured perturbation") into a veto ("disqualifying"), which would have taught the next reader to reject useful instruments. Correct AT the claim, and check the correction for the same defect. Related: [[a-detector-must-pass-a-control-built-from-what-motivated-it]], [[proving-two-ci-arms-populations-are-comparable]], [[compare-at-one-layer-crlf-meter-class]].