Slop review of the suite's own skills: remaining trims and suite-level findings

[rookslog]

Status: the de-slop and compressed-deliberation findings from this review were executed in PR #123 (merged, ee9ec68), including the P1 follow-up from its review round. This issue carries the remainder.

One measured fact changes the cost calculus for every trim below. At review time, grep -E '^\| G-' grounding/*/*.md | grep -c unaudited returned 89 of 89. Every G row in the suite read unaudited, and every E and N row leaves its audit cells empty by rule. So no cut proposed here voids a recorded audit. The costs are row deletion and renumbering. Word counts below are measured with wc -w. Every trimmable estimate and meta split is conjecture, labelled inline, from one reviewer's judgment of which sections serve a writer at invocation time.

Remaining per-file trims

proportionate-execution — 778 words, meta ≈ 283 (36%, conjecture), trimmable ≈ 185 (conjecture)

1. MAJOR — Two adjacent sections, ## How this differs from compressed-deliberation and ## How this differs from de-slop, both close "Follow both. They do not disagree." The same structural claim twice in consecutive sections, in the file that governs saying a thing once. Collapse to one section or one line. Cost ≈ 4 rows.
2. MAJOR — All five What this corrects bullets restate compressed-deliberation's vendor observations. See the duplication finding below.
3. MINOR — "A reader who can stop you needs to know what you are about to do, and a rule that removes that narration costs the reader more than it saves." A justification paragraph ahead of three bullets that already state the rule. Cut it, or cut the preceding "Some narration is load-bearing." which it duplicates.

navigable-references — 515 words, meta ≈ 184 (36%, conjecture), trimmable ≈ 140 (conjecture)

The medium/form/why table is the best single artifact in the suite. The meta around it is the second-worst ratio.

1. MAJOR — "That is a writing defect, and not a gap in the tools." An invented counterposition. Nobody claimed it was a tool gap. Cut. Cost 1 row.
2. MAJOR — ## What a check can see spends 45 words describing the capabilities of a check that does not exist. The operative residue is the final clause, "so every rule above is yours to apply". Reduce to one sentence. Cost ≈ 3 rows.
3. MINOR — "The form changes with the medium, and with how long the reference has to survive." Restates the table's own two columns immediately below it. Cut the sentence, keep the two failure examples. Cost 1 row.

subagent-returns — 693 words, meta ≈ 134 (19%, conjecture), trimmable ≈ 110 (conjecture)

1. MAJOR — "Work you did buys no words. A long return is not a report on effort, and a caller reading one pays for every line of it." Three restatements of the two bullets directly above. Cut, or keep only the cost clause. Cost 2 N rows.
2. MINOR — ## What stands behind these rules and ## Boundary state the no-source position twice, about 40 words apart. Merge, as PR refactor: two craft skills stop explaining themselves to the reader #123 did for de-slop.
3. Defended, not a finding — the "A caller re-runs a measured claim cheaply and re-derives an inference expensively" paragraph survives the deletion test. It supplies the only reason the provenance bullets are not arbitrary.

plain-language — 766 words, meta ≈ 142 (19%, conjecture), trimmable ≈ 60 (conjecture)

Tightest file in the suite.

1. MINOR — "The original site at plainlanguage.gov now redirects to digital.gov. The archived guidelines in the repository above are the source of record." Source-of-record bookkeeping a writer never acts on at invocation time. Move to source/standards/plain-language.md. Cost ≈ 2 N rows.
2. Defended — "Plain language is not simplified language." earns its negative definition, because the sibling skill in the same install is named Simplified Technical English.

simplified-technical-english — 793 words, meta ≈ 170 (21%, conjecture), trimmable ≈ 85 (conjecture)

1. MAJOR — Four separate statements of "this is not the standard, go check the standard", across Purpose, Reference routing, Final check, and Source and boundary. The most-repeated single proposition in any file in the suite, twice in near-identical wording. Keep the Final-check instance, where a writer acts on it, and one in Source and boundary. Cut the other two. Cost ≈ 3 rows.
2. MINOR — Six lines run past the 80-column wrap every other SKILL.md holds. Rewrap. Cost zero rows if no unit text changes. Verify against npm run check:ground before claiming that.

Suite-level findings

1. MAJOR — 311 words of hand-maintained sibling comparison across five craft files, six occurrences of "They do not disagree." An O(n squared) maintenance surface: a sixth craft skill obliges edits in five files. Two things must survive any consolidation: de-slop's repetition-disagreement paragraph, and each file's one-line scope claim. Constraint: no install pathway ships a suite-level document, so the alternative is one line per file, not zero.
2. MAJOR — compressed-deliberation and proportionate-execution duplicated their vendor-description sections nearly verbatim. Twelve G rows across two matrices carried seven distinct propositions. PR refactor: two craft skills stop explaining themselves to the reader #123 resolved the CD side by moving its bullets to the source record. The PE side still carries five bullets restating the same propositions. One file should state them and the other cite it.
3. MINOR — The row-taxonomy explanation and the "It is not installed with this skill" disclaimer appear in all seven files, roughly 280 words suite-wide of identical audit-record boilerplate. Defensible as honesty discipline. Priced here, not condemned.
4. BLOCKER — The shipped skills get no recorded editorial reading, and the stated warrant for the exemption does not hold. scripts/check-editorial.mjs excludes skills/ because "a grounding matrix already disposes of every unit in one, and that record has its own audit column." But that column certifies a G row against its source, never the prose against de-slop. Three matrices now have zero G rows, so no row carries an audit cell at all, and every G row elsewhere read unaudited at review time. README.md and AGENTS.md carry stamps for a human reading while the product does not. Either widen the governed list to the seven SKILL.md files, or correct the comment's warrant.

Not assessed

The two references/*.md files under simplified-technical-english, the agents/openai.yaml files, the front-matter description fields as trigger surfaces, and whether any proposed cut passes npm run check:ground or npm run lint:docs. The review ran read-only. The grounding matrices were read as the cost ledger, not as the review target.