editio

name	editio
description	Academic-writing orchestrator — turn a promptus store's validated knowledge into a defensible, submittable paper. Use when starting, resuming, structuring, rendering, or auditing a paper, or deciding which editio piece does a job. Knows the paper workspace, markdown-is-truth, three renders, the DoCO/DEO structure gate, and the evidence-calibrated audit loop.

editio — the paper read-port

Portable path rule: in commands below, replace <plugin-root> with the absolute plugin root two directories above this SKILL.md. Resolve it from the loaded skill path; do not assume a host-specific environment variable exists in the project shell.

editio turns what the store has validated into a paper that can defend itself: content is authored as per-section markdown under .editio/paper/, rendered to LaTeX, and built in three modes from one source. Every checkable claim on the page traces to a store unit at a known status, and the draft render makes that status visible before a reviewer sees it.

Prerequisite: the promptus plugin (a .promptus/ store in the host repo). editio reuses it at the skill level — recall grounds claims, the grounded-writing-reviewer audits drafts. A TeX distribution is the user's own (see editio-latex for the 5-minute setup).

Decision table — intent → do this

You want to…	Use
start or resume a paper	read and execute <plugin-root>/commands/editio.md (the /editio adapter in Claude Code)
lay down / rebuild the workspace for a venue	bun "<plugin-root>/scripts/editio-scaffold.ts" --venue arxiv
frame the argument, seed and write sections	the editio-structure skill
render markdown sections to LaTeX	bun "<plugin-root>/scripts/editio-render.ts" --all
build the PDF, preview one section, set up TeX, notation	the editio-latex skill
ground a claim before it hits the page	promptus's recall (kb-find → kb-get)
see where the paper stands (per-section claim tallies, drafted words vs budget:, grounds health)	bun "<plugin-root>/scripts/editio-status.ts" (--claims lists each ungraded/unsourced span at file:line)
audit a draft's claims + AI tells	the grounded-writing-reviewer skill, then apply its grades (below)
run the publish gate (no ungraded / unsourced / overclaims)	bun "<plugin-root>/scripts/editio-status.ts" --gate (exit 1 on violations)
check the workspace against the installed plugin (stale scaffold, declared-order drift, venue drift, hand-finished metadata, unrendered/unwired sections, identity in prose, stray PDFs in the source root, gitignored/untracked sources, builds over the venue page limit, naked file paths in prose)	bun "<plugin-root>/scripts/editio-doctor.ts" (report-only; --strict exits 1 for CI)
change the title / authors / corresponding author everywhere at once	edit paper.json, then bun "<plugin-root>/scripts/editio-identity.ts" (or editio-render --all) — regenerates front/identity.tex, the data macros every document assembles from
bind a result value once, reference it everywhere (@num:handle)	the editio-numbers skill + numbers.json; bun "<plugin-root>/scripts/editio-numbers.ts" --write
verify the paper's numbers still match their sources	bun "<plugin-root>/scripts/editio-numbers.ts" (--gate exits 1 on unknown/stale/unwritten bindings)
fix the voice / de-AI a passage	the humanizer skill
design, size, caption, or color a figure	the editio-figures skill (claim-first; venue widths from venue.json)
verify a figure PDF is the slot size	bun "<plugin-root>/scripts/editio-figcheck.ts" <fig.pdf> --slot single
tables, bibliography, venue packaging, rebuttal	editio-tables / editio-bib / editio-venue / editio-rebuttal (later phases; not yet shipped)

The invariant (inherited from promptus, applied to manuscripts)

Markdown is the only source of truth — the .tex siblings are derived and disposable (regenerate any time; the header comment says so). Evidence truth stays in the store. The renderer never blocks: ungraded and unsourced claims render (tinted in draft); enforcement — zero .unsourced, no overclaims — is a lint gate at publish time, not a render failure. Identity lives in paper.json only, scaffolded as placeholders — and it reaches the documents as generated data macros (front/identity.tex, via editio-identity), never hand-written: one paper.json edit updates the title, author block, and bios everywhere.

The audit loop (the hinge)

1. Draft — write prose in sections/<slug>.md; wrap checkable claims in spans: [the gate refuses off-vocab writes]{.claim} (ungraded is fine — grading comes next).
2. Retrieve — recall looks each claim up (kb-find → kb-get), returns substrate:status.
3. Grade — the grounded-writing-reviewer skill (read-only) reports findings per flagged span (unsupported / over-confident / style tells); you map them to grades: finding:VALIDATED/lit:CITE → .validated · CONJECTURED/provisional → .conjectured · nothing found → .unsourced · DEADEND/REFUTED backing the claim → an overclaim flag.
4. Apply + override — you (the session) write the grades back into the spans, adding grounds=<handle>; the author accepts, overrides in-span with a reason (override="holds for our corpus"), or fixes the prose / stores the evidence.
5. Render — grades become \claimV (clean) / \claimC (amber) / \claimU (vermilion) / \claimG (grey, ungraded) on the draft page.
6. Gate — editio-status --gate: zero overclaims (a .validated claim over weak, unknown, or absent grounds), nothing left ungraded, and every .unsourced claim either fixed or explicitly accepted. override="reason" passes the gate on the record exactly twice over: on an .unsourced claim (the author's acceptance, printed in the gate output) and on a .validated claim's weak/unknown grounds — never on an ungraded span (ungraded means the loop hasn't run; the fix is running it, not excusing it). publish/blind strip every tint either way.

Three renders, one source

main.tex defaults \editiomode to draft; publish collapses annotations; blind additionally masks \selfcite, drops \blindhide{…} blocks, and swaps the author block for the anonymous label from paper.json. Venue is orthogonal (a class + data swap).

When NOT to use editio

• No store behind the claims → it's typesetting, not grounded writing; build the store first (or accept every claim staying .unsourced).
• A one-off memo or README → plain writing + humanizer; the paper machinery earns nothing.