remora-deslop

name	remora-deslop
description	Use when writing or rewriting any reader-facing Markdown in REMORA-research (README, docs/, ARCHITECTURE, paper .md), when the prose-style ratchet (scripts/check_prose_style.py) fails, or when asked to "deslop", "fjern AI-slop", "rydd prosa", "skriv om dokumentasjon". Removes structural AI-writing tells without touching claims, numbers, anchors or preserved history.

remora-deslop

Adapted from stephenturner/skill-deslop and poteto/noodle unslop. The vocabulary catalogs in those skills were measured at zero in this repo (2026-08-26). What this repo has is structural: em dashes at about five per thousand words, arrows inside sentences, bold-first bullet lists. This skill targets those, and adds the constraints an assurance repository needs that neither source has.

Hard constraints (never violated by a prose pass)

1. Numbers, percentages, counts, dates, commit SHAs, issue numbers, CLAIM-/ CAP-/REM-/RF-/DOC- identifiers and claim anchors are copied verbatim. A prose pass changes wording only. If a number looks wrong, stop and report it; do not fix it in passing.
2. Generated documents are rewritten in their generator, never in the output. Register status generated and generated_by name the script.
3. NEGATIVE_RESULTS.md section markers (open/accepted/superseded) and any text labelled superseded, historical or withdrawn are preserved word for word. docs/archive/, attic/ and docs/researchpapers/ are out of scope entirely.
4. Pre-registered and frozen documents are never reworded after the fact: docs/assurance/statistical_analysis_plan_v*.md, superseded_claims.md, experiments/*/PREREGISTERED.md. Their counts simply stay in the baseline.
5. The paper is edited in paper/remora_paper.md and paper/remora_paper.tex together; the PDF is master-owned. scripts/check_paper_sync.py must pass.
6. README.md stays within its line budget (check_document_governance.py).
7. Every gate that touches text must pass before the PR: document governance, links, readme-claims, paper-sync, claim-provenance, and scripts/check_prose_style.py (counts may only fall; then run --update-baseline and commit the baseline in the same PR).

What to remove

• Em dash in prose. Rewrite the sentence: split it, use a comma, a colon before a list, or parentheses. Do not search-and-replace the character.
• Arrow (→, ⇒) inside a sentence. Write the relation in words. Keep arrows in pipeline notation lines (A → B → C), tables, code and diagrams.
• Bold-first bullets (- **Label:** text). Turn the list into prose, or a table if the items really are label/value pairs, or plain bullets whose first words carry the meaning without bold.
• "Not X, but Y" / "not just X, it's Y". State Y.
• Filler openers and transitions: "it is worth noting", "it should be noted", "importantly,", "notably,", "in order to".
• Superficial participle tails: ", ensuring safety", ", highlighting the need for". Either say concretely what is ensured, or delete.
• Copula dodges: "serves as", "stands as". Use "is".
• Tool or assistant residue in reader-facing text: "ChatGPT analysis", "as the assistant noted", "let me know", "hope this helps". Attribution of AI assistance belongs in docs/AI_USE.md only.
• Triple stacks of short fragments for emphasis; self-posed questions answered in the next sentence; "In conclusion"; "Despite these challenges".

What to keep (rejected from the source skills)

• Passive voice and impersonal register where the subject is the system or the evidence. "The grant is consumed atomically" is correct technical prose.
• Lists of three when there are three things. No tricolon ban.
• Precise domain terms (fail-closed, nonce, envelope hash, PEP, PDP).
• Caveats and hedges that carry information ("not claimed as production evidence"). Remove only hedges that stack without adding a condition.
• No "soul", opinions, first person or deliberate mess. Assurance documents are read by reviewers looking for claims; voice is not the goal.

Lessons from the 2026-08-26 pass (six layers, 1160 -> 298 em dashes)

• Never touch headings. The scanner skips them and tests key on heading text (## RF-10 in docs/13 broke test_research_shelf). Restore any heading the rewrite touched from origin/master before committing.
• Condition-to-outcome bullet lists (- X -> Y) are tables, not prose. Convert them; tables are exempt from the scan by design.
• Label/value bullet schemas (Claim/Evidence/Artifact, Purpose/Storage/Gap) keep their structure; only the bold comes off the label.
• Generated documents: change the generator's templates, run it, and make sure --check passes; never edit the output.
• A dash at end of line becomes an orphan ; under mechanical joins; always sweep for ;$ and for **Label**; and ``code; joins, which read better as a colon or comma.
• Frozen by policy and excluded from the scan: register-historical documents, pre-registered SAPs and protocols, results/, data/, datasets/, tests/ fixtures, the credibility pack, CHANGELOG.md. Still open for a deliberate decision: NEGATIVE_RESULTS.md (a permanent record) and the paper (.md and .tex in lockstep, PDF master-owned).

Process for a systematic pass

1. python scripts/check_prose_style.py --report to pick the next file by count. Work one layer per PR: reader-facing entry docs first (README, ARCHITECTURE, DEVELOPER_OVERVIEW, CONTRIBUTING, docs/00-13), then docs/assurance, then docs/design and roadmap, then generators, then paper.
2. Read the whole file before editing. Rewrite sentence by sentence; keep each paragraph's claims and their order.
3. Run the gates in constraint 7. Diff the file and re-read the diff for any changed number or identifier.
4. python scripts/check_prose_style.py --update-baseline, commit the baseline with the text.
5. One PR per layer, titled docs(prose): <layer>; the PR body lists the before/after counts from the scanner.