---
name: stating-a-limit-does-not-route-around-it
description: "I named my instrument's inconclusive side and its deletion blind spot, then handed over a verdict that depended on both — naming a limit is not the same as routing around it."
metadata: 
  node_type: memory
  type: feedback
  originSessionId: c63f1093-4665-42cb-a581-0f78a683e349
  modified: 2026-08-22T08:52:53.590Z
---

Measured 2026-08-22, w3-lane assembly. I built a content-absorption instrument (is every
substantive line this commit ADDS present anywhere in the head tree?) and reported its
asymmetry unprompted, in the same message as the verdicts:

- **zero residue = trustworthy** (the added text is all there)
- **residue > 0 = INCONCLUSIVE** — either genuinely needed OR the head evolved past it, and
  *only reading the residue tells you which*
- it **cannot see DELETIONS**, so a removal-shaped commit reads "absorbed" by construction

Then, in that same message: I read the residue for `43dc3943`, caught that the head's
paragraph was 145 words LONGER (rename plus a later amendment) and correctly called it
absorbed-and-superseded. And for `c9544f71` I did **not** read the residue. Six lines across
three files, containing a REQ title and a cell name, *looked* substantive at a glance, so I
wrote "genuinely needed, confirmed independently". It was absorbed. The deciding evidence was
a DELETION — a struck phrase surviving at the head as a comment saying it is struck and a test
asserting it must not return — i.e. exactly the blind spot I had named one paragraph earlier.

**The failure was not the limit. It was applying my own stated rule to one commit and skipping
it on the next**, because the second one's residue matched the answer I already expected. The
tell was sitting in the residue: it contained a **REQ TITLE**, and a requirement title
differing between lane and head is the single most likely thing to be a reword — which is
precisely what the commit I *did* read turned out to be.

**How to apply:** when an instrument has an inconclusive side, the inconclusive verdict is not
a verdict — it is a work item, and it must be discharged per item, never per batch. Write the
open one down as OPEN. *Three settled and one open beats four settled and one wrong*, and the
difference here was one paragraph of work I chose not to do. And when the deciding evidence
for a call falls inside a blind spot you have already named, that is the moment to route around
it deliberately rather than to trust the reading — naming a blind spot does not stop you
walking into it.

**What caught it:** a `git cherry-pick` CONFLICT, which forced a human read of both sides. Of
the five wrong populations produced across two agents that day, the four others were caught
because someone chose to re-measure; this one was caught because a mechanical process
**REFUSED**. A refusal is the only failure mode in that day that did not depend on somebody
deciding to look again — so prefer instruments that can refuse over instruments that can only
report.

Related: [[per-line-suppression-resumes-inside-a-multi-line-token]] (a red that did not fire),
[[dont-take-a-diagnosis-as-measured]], [[census-enforcers-and-staters-of-an-invariant]],
[[read-source-at-the-measured-sha-not-the-default-tree]].