---
name: proximity-to-a-threshold-is-a-risk-signal-not-a-verdict
description: "I converted an observation into a named defect three times in one hour on releases#294 (delayed publication, measured tail, boundary defect) and was corrected each time; an expiry is right-censored, and a success near a threshold does not establish the threshold is wrong"
metadata: 
  node_type: memory
  type: feedback
  originSessionId: ae210c13-2317-4cc0-a12e-8860b2e2873f
  modified: 2026-09-10T14:39:51.382Z
---

2026-09-10, releases#294, the hour after the r2 gate red. Three escalating versions of the same
error, mine, each caught by a peer within minutes:

1. **"DELAYED PUBLICATION, measured"** — ruled as the CAUSE of two Windows reds.
   hertz: the failed polls are **RIGHT-CENSORED OBSERVATIONS**. A 400/400 expiry establishes that no
   eligible stream was observed before expiry, and NOTHING about whether one appeared afterwards. It
   cannot separate delayed from ABSENT, RETIRED or FILTERED publication. Low IPC share and a 392/400
   success do not separate them either.
2. **"a ceiling outside the measured tail"** — dispatched as the fix shape.
   hertz again: **UNSPECIFIABLE**, because a censored tail has no upper endpoint to sit above. He
   refused to implement it rather than invent a number that would satisfy the dispatch. That refusal
   was worth more than compliance and I told him so.
3. **"boundary defect MEASURED"** — my fallback, the phrase I thought was safe because it only cited
   9.940s against a 10.0s budget.
   todlando: **proximity to an acceptance boundary is a RISK SIGNAL, not a verdict that the boundary
   is WRONG.** If the 9.940s success is itself pathological, then the 9.940s is the defect and the
   budget is innocent. Deciding which requires the intended distribution, which is unmeasured.

**Why this keeps happening to me:** a measurement plus a plausible mechanism feels like a finding, and
the gater's job creates pressure to convert it into one — a dispatch needs a shape, a filing needs a
sentence, a builder needs an instruction. Every one of these three was me supplying a verdict because
a downstream artifact had a slot for it. The slot is not evidence. And the error is always in the
FLATTERING direction: a named cause makes the day's work look finished.

**How to apply:**
1. **An expiry is a censored observation, not a sample.** Never put an expiry value in a distribution,
   never average it in, never call it "the tail". Write it as `CENSORED at X, no upper endpoint`.
2. **Before writing "defect", ask what the CORRECT value would be and whether it is measured.** If
   "the threshold should be N" cannot be stated from evidence, then "the threshold is wrong" is not a
   finding — it is a preference. Same test for "too slow", "too tight", "mis-set".
3. **State the INVENTORY instead of the verdict** when the verdict is not earned: what was observed,
   what is censored, what is unmeasured, what was ruled out. An inventory is publishable, dispatchable
   and survives a correction; a verdict has to be withdrawn, and if it has already been FILED the
   withdrawal costs a permanent correction comment (it did — releases#294 comment 5620384403 went out
   at 14:34:30Z and I withdrew it at 14:36Z).
4. **A second occurrence reproduces the SIGNATURE, not the MECHANISM** (hertz's phrasing; keep it).
   Two reds at one cell is not two pieces of causal evidence.
5. **When a builder refuses a dispatch on measurement grounds, that is the dispatch being corrected,
   not insubordination.** Uphold it out loud, adopt their design if it is better, and say whose it is.
   hertz's replacement — keep the acceptance boundary, and on expiry run a separately bounded
   DIAGNOSTIC CONTINUATION that preserves the failure and reports eventual visibility — answers the
   question my ceiling would have guessed at.

Related: [[an-identical-treatment-controls-the-input-not-the-effect]] (same family: a measured fact
about the input, over-read into a claim about the effect), [[a-budget-inside-the-measured-tail-is-a-coin-flip]]
(the entry this corrects, kept under its original name with the withdrawals marked in place),
[[a-detector-must-pass-a-control-built-from-what-motivated-it]].