---
name: identical-readings-across-opposite-outcomes-indict-the-meter
description: When a control and a treatment produce the SAME reading, that is evidence about the instrument, not a finding about the subject — and it is the strongest such evidence you get.
metadata:
  type: feedback
---

When two runs you expect to DIFFER report the same number, the first hypothesis is that the meter is
broken — not that you have discovered the two cases are secretly alike. A control and a treatment
agreeing perfectly is the cheapest, loudest instrument check available, and it arrives for free
inside work you were already doing.

**Why:** measured 2026-08-21, from my own #153. On 2026-08-04 I filed "a structural refusal from
`spt update --restart` exits 0" off a captured `EXIT=0`. The same ticket ALSO reported that a
*successful* roll on a second node exited 0, and I wrote that up as corroboration — "exit 0 currently
covers both rolled and refused". It was not corroboration. It was the refutation, sitting in my own
filing as a supporting paragraph. Both runs used `cmd 2>&1 | tail -N; echo "EXIT=$?"`, so both
measured `tail`, which exits 0 unconditionally; spt returns 3 and always did, at four trees. The
issue was wrong, sat 17 days, and ate the front of a builder's lane before todlando asked how the
number was captured. See [[exit-code-after-a-pipe-is-the-tails]] for the concrete trap.

The failure is not "I did not check the instrument." It is that the identical reading was in front of
me, already written down, and I had classified it as a SECOND OBSERVATION supporting the finding
rather than as a contradiction of the method that produced both. Two data points agreeing is
persuasive when they are independent; when they share a wrapper, they are one observation wearing two
coats — the same shape as [[a duplicate summary row is not a second observation]] and the per-HOST
occurrence check.

**How to apply:**
- Before quoting a measurement, ask what reading the OPPOSITE outcome would have produced under the
  same capture. If the answer is "the same one", you have measured your harness.
- Treat a control that agrees with the treatment as a red flag that outranks the finding. Stop and
  vary the instrument, not the subject.
- Deliberately include a case you expect to FAIL when building any capture you intend to quote — a
  meter that cannot produce two different answers is not yet known to produce one.
- Distrust "and it reproduced on another node" when the other node ran the same wrapper. Independence
  is a property of the METHOD, not of the hosts.

Related: [[dont-take-a-diagnosis-as-measured]] (this is that rule's instrument-side face),
[[exit-code-after-a-pipe-is-the-tails]] (the trap that produced it),
[[relay-is-not-the-gaters-word]] (same session, the other half of the day's measurement lessons).