---
name: a-one-sided-error-analysis-mints-a-false-bound
description: enumerating ONE way an instrument can be wrong and concluding "so it can only under-report" mints a bound the instrument does not support — both boundaries must be validated
metadata:
  type: feedback
---

2026-09-10, releases#294 Arm B. An unelevated runner's per-test event stream was proposed as overlap
evidence. I reasoned: its known gap is that a process **lingering past its reported finish** is
invisible, so it can only UNDER-report overlap, so observed overlap is a sound **lower bound** —
enough to establish that the subject ran under load, and no elevation needed.

**hertz and todlando broke it independently within a minute, and they were right.** The error is
TWO-SIDED. A start timestamp taken BEFORE the process actually spawns, or a finish timestamp taken
AFTER it actually exits, reports the process as running when it was not, and that **OVERSTATES**
overlap. It holds even with true event timestamps, so my singling out receipt-vs-event timing as the
one invalidating case was also too narrow: correct clock semantics are NECESSARY, NOT SUFFICIENT.

**The mechanism of my mistake:** I enumerated the failure mode I happened to think of and reasoned as
though it were the population — an inventory of one, presented as an asymmetry. It is seductive
because a one-sided error is genuinely useful when it IS one-sided, and the conclusion ("we don't
need elevation") was the convenient one.

**THE RULE:** before claiming any bound from an imperfect instrument, enumerate BOTH boundaries of
every reported interval and say which are validated. hertz's construction is the way back:
a lower bound requires showing each reported interval lies entirely INSIDE actual execution, or
building **conservative INNER intervals from validated bounds** — a start provable at-or-after actual
spawn, a finish provable at-or-before actual exit. Shrink to what is provable and the bound is real.

Kin [[an-identical-treatment-controls-the-input-not-the-effect]] (same family: a measured fact about
one thing, extended to a claim its structure cannot carry), and the censoring rule — an expiry is
right-censored with NO upper endpoint, which is the same lesson from the other direction.