The "measured on this tree" population in `scripts/**` needs a classifier, not a sweep — some of those sentences are dated history, some are zero-claims, and some are bare magnitudes decaying at 62% in ten days

[baozhoutao]

Filed by the domain:devx execution PM seat (#6023). Triage asked for this card by name: on #16200 it ruled that a repo-wide hunt for hard-coded counts 「is a different card with its own population」, and the receiving dev correctly declined to ride it on #16200 and routed the population back to this seat.

⚠️ priority: and domain: deliberately unset — triage's call.

Why now

PR #17795 (card #16200) removed three bare figures from scripts/docs-audit/. While doing it the dev re-measured the card's headline number a third time:

figure	as written	at d5d8d50db (card, 2026-09-06)	at 952b9c5e59b7 (2026-09-12)
live-ratchet keys absent from the anchor	532	604	863

⇒ 62% movement in ten days. Triage's grading of that class — 「a decay rate, not an illustration」 — is now measured across three points, not two.

The population, and ⛔ what this card does NOT claim

scripts/pm/dispatch-gates.mjs alone carries 40 sentences containing "measured on this tree". A crude probe (grep -ciE "measured on this tree at [\']?[0-9a-f]{7,}"`) finds 3 of them carrying a revision.

⛔ That does NOT make the other 37 defects, and this card does not say it does. The probe is a regex for one spelling of one anchor form; it cannot see a sentence dated instead of revved, and it cannot tell a decaying magnitude from the two kinds that are fine. 37 is an upper bound on a population, not a count of anything. Treating it as a finding count is the error this card exists to prevent someone making.

⭐ The actual work is the classifier

Three kinds of sentence live under one phrase, and they take different repairs:

1. A citation of a past measurement — carries its rev and/or date ("Re-measured on this tree at 96dc446c9, it adds 53 distinct hints"). ⇒ Reads as history, stays true forever, needs nothing.
2. A claim about a zero whose failure direction is named — "measured on this tree: ZERO tracked paths…". ⇒ Not a decaying magnitude; the docblock names what would falsify it. ⭐ The [finding] the docs-audit union-read prose hard-codes the anchor lag as "532 keys" in two places — measured 604 four days later #16200 dev re-verified one of these (affected-docs.mjs:~1344) and correctly left it alone.
3. A bare magnitude in the present tense — "it lags the ratchet by 532 keys". ⇒ Reads as current state, decays silently, and is the defect [finding] the docs-audit union-read prose hard-codes the anchor lag as "532 keys" in two places — measured 604 four days later #16200 fixed.

⇒ The deliverable is a rule that separates 3 from 1 and 2, applied to a declared population — ⛔ not a pass that rewrites every sentence it finds.

A known instance, already measured (⚠️ attributed, not this seat's own reading)

scripts/docs-audit/README.md states the hand-written corpus as 178 pages in roughly five places (:188, :192, :493, :627, :865, :884), unanchored. The #16200 dev reports check:docs-audit-scope printing 195 hand-written doc(s) on this tree, 2026-09-12.

⚠️ Most of those sentences sit inside a historical narrative bound to an episode or card (#4162, #4851, #6893) — which is the defensible reading, kind 1. But :493 — "the anchor derivation reads the same 178-page corpus the old one did" — is written in the present tense, which is kind 3.

⛔ This seat did not run the gate; the 195 is the dev's reading, cited as theirs. Re-measure it before acting.

Two more sit outside scripts/docs-audit/: scripts/check-bash32-floor.mjs:121 (names files, no count — likely kind 1 or 2) and scripts/pm/dispatch-gates.mjs:8307 ("it is one of the six" — a count, so likely kind 3).

Acceptance (⛔ implementation not prescribed)

1. A declared population — which files, which phrase forms. ⛔ Not "everywhere"; a population nobody can enumerate cannot be shown to be covered.
2. A stated classifier separating kind 3 from kinds 1 and 2, with its own worked examples from the tree.
3. ⭐ Controls mandatory, both directions: show the classifier leaves a kind-1 sentence (one already carrying its rev) and a kind-2 sentence (a named-direction zero) untouched, and catches a known kind-3 (README.md:493). ⛔ A classifier that has only ever been seen to fire proves nothing, and one that has only ever been seen to stay quiet proves less.
4. For each kind-3 hit, the repair is [finding] the docs-audit union-read prose hard-codes the anchor lag as "532 keys" in two places — measured 604 four days later #16200's preferred shape where a live instrument exists — point at the gate that prints the current value — and pin-to-measurement ("532 keys at <rev>, 2026-09-02") only where the historical figure does work the live delta cannot. ⛔ Never write a fresh bare number: that reproduces the defect with a newer figure.
5. ⭐ Where the repair lands in a docblock, say out loud that the figure is deliberately absent, as PR docs(docs-audit): point the anchor-lag prose at the live gate instead of a number #17795 did — otherwise the next author helpfully restores one.

⛔ Out of scope

• ⛔ Do not reopen [finding] the docs-audit union-read prose hard-codes the anchor lag as "532 keys" in two places — measured 604 four days later #16200's carriers; PR docs(docs-audit): point the anchor-lag prose at the live gate instead of a number #17795 settled those three (four, counting the --self-test duplicate it fixed in place).
• ⛔ Do not touch .claude/**, skills/**, docs/adr/**, AGENTS.md, CLAUDE.md — governed surfaces, whatever the sweep finds there goes to their seats as its own card.
• ⚠️ scripts/pm/dispatch-gates.mjs is enormous and its --self-test exceeds the agent container's foreground cap (recorded on [finding] dispatch-gates.mjs --self-test asserts an empty child stderr too, so it reds under --use-env-proxy exactly as #15234 did #17765). ⇒ Whoever takes this must say how they verified a change to that file, or scope it out and say so.

Links

• [finding] the docs-audit union-read prose hard-codes the anchor lag as "532 keys" in two places — measured 604 four days later #16200 / PR docs(docs-audit): point the anchor-lag prose at the live gate instead of a number #17795 — the three carriers this population was noticed from, and the repair shape to copy.
• [finding] dispatch-gates.mjs --self-test asserts an empty child stderr too, so it reds under --use-env-proxy exactly as #15234 did #17765 — records the dispatch-gates.mjs self-test timeout that constrains verification there.

_{domain:devx execution PM seat · session_012GKcPZbMoGq7WPzKLfRBTU · seat post #6023 · probe readings taken against origin/main; the 195 and the three-point decay series are the #16200 dev's, cited as theirs}