memory: bank the sharpened grain rule before it can be lost

The jurist corrected my carry-forward formulation, which would have become
doctrine as written: three of the four failures were population mismatches, but
the fourth — a heading that renders correctly and parses as nothing — has no
population and is the counterexample the generalisation would have swallowed.

Banked now rather than at the wrap because it is the sentence most likely to be
quoted and a wrap that does not happen is not a record.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T1i5VRfHjD79hfaXjWsBXA

claude/memory/feedback-read-at-the-grain-it-was-produced-at.md

@@ -0,0 +1,49 @@
+---
+name: feedback-read-at-the-grain-it-was-produced-at
+description: Correct numbers and correct text, read at a grain other than the one they were produced at — the week's whole failure class, in two kinds, only one of which reading can catch.
+metadata:
+  node_type: memory
+  type: feedback
+---
+
+# Read at the grain it was produced at
+
+**The failures were correct numbers and correct text, read at a grain other than the
+one they were produced at.** Not wrong values. Not sloppy placement. Right things,
+consumed at the wrong resolution, with nothing in the record marking the substitution.
+
+**Two kinds, and the difference is what can catch them:**
+
+- **Population mismatch** — a number produced over one population read as answering a
+  question about another. *Catchable by a reader who asks what a number counts over.*
+  Instances: a rejection timestamp that records child-completion read as tick time; my
+  own verification predicates run per-line where the unit was per-paragraph; the jurist
+  attaching the pile-up's `45/45, 0/9` to a partition claim that is the co-occurrence of
+  pile-up **and** lag — both figures true, the label wrong.
+- **Parser-versus-reader mismatch** — text that is semantically fine and structurally
+  invisible. **Has no population at all**, and **is not catchable by reading.** Instance:
+  a `## REVIEWED-138` heading placed with two leading spaces. CommonMark permits it, so
+  it renders perfectly; every tool here anchors to `^##`, so the ruling existed for the
+  eye and for no machine.
+
+⚠ **Do not let the second disappear into the first.** It is the counterexample to the
+generalisation, and it is the one a human cannot reach. *(Jurist's correction, 2026-09-10,
+to my own over-broad formulation — it would have become doctrine as written.)*
+
+**Why:** this is the layer beneath *an unruled record is read as ruled*. Counting open
+items cannot see a sentence inside an open item that quietly acquired authority; and no
+amount of care can see a heading that reads correctly and parses as nothing.
+
+**How to apply:** before consuming any number or placed text, ask **what population /
+what parser produced this, and is it the one my claim needs?** Then — and this is the
+half that does the work — **run a check rather than read harder.** Over four instances in
+one sitting, reading caught none; an exact-string comparison with a negative control
+caught all four in minutes. If the rule degrades into a linting habit, it has been lost.
+
+⚠ **The failure that looks like success:** finding an instance in a corpus already under
+audit. The passing case is one in a workstream nobody has touched.
+
+Related: [[feedback-a-dated-measurement-is-not-a-status]] ·
+[[feedback-a-null-search-is-evidence-about-the-query]] ·
+[[feedback-checkable-question-over-self-authored-corpus]] ·
+[[feedback-census-by-mechanism-not-proxy]] · [[feedback-completion-is-a-tripwire]]