site: copy edit — 7 pieces tightened

daily-finds.log.md

@@ -5441,3 +5441,30 @@ Two, both mine rather than the copy editor's:
 - **backscatter-radar-site** and **madrid-model-1830** were four and five on the shortlist and both would have made good reviews. Held on source access rather than merit: the radar site has no first-party site of any kind (it is an abandoned military installation), and the Madrid model's museum page is thin enough that the fact-check would have been a second pass over the same Atlas Obscura text. Both stay strong ungraduated candidates.
 - **catalogue-of-ships** was the most tempting hold of the run — it is seventh on the shortlist and its subject is exactly right for a long review. It has no hero and there is no image on the domain to take, and the `the-listening-museum` decision on 08-07 (a title card is not a hero) says a review published with nothing to look at is worse than one that waits. It waits.
 - **the-phone-in-your-hand** remains ungraduated for the second run running, still on the same OG-card hero gap flagged on 08-07. Still the strongest ungraduated candidate on the board by story.
+
+## 2026-08-08 — copy edit
+
+Fresh-eyes sonnet subagent over all 41 pieces written this run (31 find bodies, 7 bundle blurb-plus-intros, 3 review bodies). **Seven files changed, and for the first time the pattern was in the conductor's own copy rather than the writers'.**
+
+**The finding.** Six of the seven bundle intros ran three sentences against the charter's 1–2 sentence budget, and in every case the third sentence was the same failure: a closing thesis restating the blurb sitting directly above it. `the-thing-is-gone-the-shape-stayed` said "What survives here is never the thing itself — it's the impression it left"; the blurb one line up already said "Each of these is the negative space where something used to be." That is principle 3 exactly ("no restating in the close what the lead already said"), committed six times in a row by one author in one sitting.
+
+The edits, all accepted on review:
+- `go-for-the-one-strange-fact` 51→31 words
+- `made-to-be-listened-to-properly` 56→40
+- `one-detail-they-wouldnt-leave-alone` 52→32
+- `the-thing-is-gone-the-shape-stayed` 52→31
+- `what-it-wont-do-is-the-feature` 50→31
+- `you-work-it-out-by-playing` 48→41
+- `stopthescript` (find) 61→45 — untangled a sentence carrying three subordinate clauses into one clean claim. No fact lost; this one is a real improvement rather than just a cut.
+
+`the-wrong-size-on-purpose` was already at two sentences and was left alone, which is the control that says the subagent was reading rather than pattern-matching.
+
+**Two cuts examined before acceptance, because each removed the sentence that named the theme.** `you-work-it-out-by-playing` lost "Nothing here explains itself in prose first," and `what-it-wont-do-is-the-feature` lost "Subtraction is the design here." In both cases the through-line survives — in the title and again in the blurb, both of which render above the intro on the page. So the reader loses nothing and the intro stops saying the same thing three times. Accepted. Had the title been vaguer, these two would have been the reverts.
+
+**The lesson worth keeping, since it is new.** Seven bundles written back to back in one pass drifted into a house formula — two concrete examples, then a summarising flourish — and the author could not see it because every instance looked fine on its own. This is the first run where the copy-edit stage earned its keep on *bundle* copy; on 08-01 through 08-07 the finding was always a long review. When a future run writes more than two or three bundles at once, the third sentence is where to look first.
+
+**Non-findings worth recording.** All 31 find bodies measured 45–85 words inside the 40–90 band and none needed a cut; the subagent reported no ad copy, no throat-clearing openings and no redundant closes across any of them. All three reviews came in clean and inside budget (`oshirado` 485 against its ~800 narrative allowance, `eigendrum` 401, `karafuneya-coffee` 366, both against 250–450). On a 41-piece day with 31 finds from five different writers, that is the capture stage writing to the charter rather than being trimmed into it afterwards — the same result as 08-07, now at 1.5× the volume.
+
+The joke-format exception held again: `joke-nick-nack` is 21 words against a 40–90 budget and the subagent was told the precedent up front (`joke-eggs-break-up` 10 words on 08-07, `joke-cashew` 13 on 08-03, `joke-ground-beef` 11 on 08-02) and correctly left it untouched.
+
+**The counting-convention drift did not recur, and that is because the method changed.** The subagent counted with a script that strips frontmatter, imports and alt text rather than eyeballing, and its numbers matched the main session's on every file. Four previous runs (08-02, 08-03, 08-04, 08-07) logged a consistent ~12-word overcount from em-dash and hyphenated-compound tokenisation. Scripted counting closes it. Not worth re-deriving a sixth time — worth adopting.