scriptorium v0.3.0

The v0.3 release. Four skills ship: reporting-guideline-compliance,
figure-text-alignment (text-only), compression and
outlier-sentence-detector, for sixteen in total. Every note in
knowledge/ was re-verified against its sources in the October 2026
knowledge refresh, and the skills were updated to match. The author
now decides what skills may do and where: policy and scope guards are
defaults the author can override for one run or in meta.overrides
in MANUSCRIPT_STATE.yaml, and accuracy guards are fixed. The skills
can also be installed with npx skills add seandavi/scriptorium for
any agent the skills CLI supports, and the CLI runs on Python 3.14.
scriptorium prompt-pack now writes one file per skill by default
(see Changed).

Test count: 226 (was 114 at v0.2.0).

Not in v0.3: statistics-consistency, voice-profile and
persona-calibration. Each has a design memo under docs/design/ and
is deferred to a later release.

Author authority

The author now decides what skills may do and where they apply (#156).
Skill refusals are declared guards of three kinds. Policy guards
(funder and publisher rules) and scope guards (the declared-work
boundary) are defaults the author can override. Accuracy guards (no
citations from model memory, no altered statistics, no inference the
input can't support) are fixed. The defaults are unchanged: skills
still decline blank-slate generation unless the author overrides it.

• Mechanism (#157, PR #160) — the
  knowledge/conventions/author-authority.md convention; a guards:
  list in skill manifests (id, kind, summary, source,
  default of ask or refuse); meta.overrides in
  MANUSCRIPT_STATE.yaml; and scriptorium validate cross-checks
  that reject an unknown guard id, an override of an accuracy guard,
  or a skill that doesn't declare the guard. A per-run answer takes
  precedence over meta.overrides, which takes precedence over the
  default. Scope and accuracy guards must default to refuse
  (enforced by tests/test_manifest_guards.py).
• Skill conversion (#158, PRs #161–#166) — 31 guards (2 policy,
  15 scope, 14 accuracy) declared across 13 skills. Each skill states
  a guard's concern and source in one line at every guidance level,
  says how to override when it declines, and names accuracy guards as
  fixed. explain, init and tour declare no guards but describe
  the mechanism; init offers meta.overrides as an opt-in field.
• "Overrides applied" output line — a skill that proceeds past any
  guard, through a per-run answer or meta.overrides, lists the guard
  ids on one line under its report title, including an ask guard the
  author answered yes to. The line is omitted otherwise.
• Editorial-side use asks instead of refusing —
  reviewer-simulation, desk-rejection-risk, venue-fit and
  author-contribution-audit share the editorial-side-use policy
  guard (default: ask). It fires when the user is reviewing or
  editing a submitted manuscript, and names NIH's outright ban.
• New opt-in capabilities, each an overridable scope guard that
  defaults to refuse:
  • verified citation suggestions in citation-audit and gap-finder
    (citation-recommendations): only works resolved against
    Crossref or PubMed in the same session are named;
  • labelled image reading in figure-text-alignment
    (figure-image-reading): every image-derived finding is labelled
    "unvalidated vision", and image-integrity claims stay barred;
  • rewrite suggestions in figure-text-alignment and
    outlier-sentence-detector (critique-rewrite-suggestions): one
    labelled suggestion per flagged item.
• Docs (#159) — a new Author authority concept page with a guard
  table generated from the manifests; meta.overrides in the schema
  reference; glossary entries for guard, override and accuracy guard;
  the declared-work-scope page, DESIGN.md, README and landing page
  describe the defaults and author authority. The roadmap resolves the
  specific-aims decision: NIH application prose is allowed behind
  the nih-application-prose policy guard (default: ask, NIH
  NOT-OD-25-132).

Knowledge refresh (October 2026)

Every note in knowledge/ was re-verified against its sources (#125).
Identifiers were checked against Crossref, PubMed, arXiv or the
publisher, misattributed findings were corrected, unverifiable figures
were removed, and 2025–2026 evidence and policy changes were added.
The shipped skills were then updated to match.

Knowledge notes

• Peer-review evidence and platforms (#136) — statistics and
  overclaiming critiques are common but usually fixable (Bordage 2001,
  Bornmann 2010); suggesting a checklist to reviewers had no effect in
  an RCT (Cobo 2007); 2026 reviewer-AI policies at Elsevier and
  Springer Nature; hidden prompt injections in manuscripts.
• Reporting guidelines, reproducibility and research integrity
  (#137) — CONSORT 2025 and SPIRIT 2025 (renumbered items), the
  LLM-era reporting guidelines (TRIPOD-LLM, REFINE, CHART, GAMER),
  LLM adherence-checking studies, TOP 2025, and a corrected Statcheck
  validity record (Schmidt 2017 is a critique) and Carlisle figure.
• Grants and significance (#138) — NIH Simplified Review Framework
  scoring (Factor 1 caps Overall Impact), the end of paylines, the
  NOT-OD-25-132 AI-originality policy and NOT-OD-23-149; NSF and HHMI
  updates; Teplitskiy et al. 2022 replaces a misattributed "Lin et al.
  2022".
• Citations and scholarly tooling (#139) — corrected Cochrane review
  denominators (medians are per journal), the reversed Sauder 2022
  finding, Greenberg 2009 denominators, and the four-step protocol
  re-attributed to the reference-accuracy / editorial spot-check
  protocol (Greenberg supplies the distortion categories).
• AI and LLMs in scholarly writing and review (#140) — the ICLR 2025
  RCT is Thakkar et al., not Liang; Kobak et al. updated to the 2025
  Science Advances version; prevalence data on hallucinated
  references (NeurIPS 2025, the Lancet audit, ICLR 2026).
• Critique techniques, editing, author roles and conventions (#141)
  — gap-finder's seven categories come from Miles (2017; Adu & Miles
  2023), not Robinson 2011; BERTScore's documented weakness is
  negation and numerical insensitivity, not "antonymy"; van Gelder's
  argument-mapping figures are pooled pre/post studies.
• Publishing, venues and editorial policy (#143) — journal-disclosed
  desk-rejection figures replace trade-press rates; Science's Board
  of Reviewing Editors breakdown (Zhang et al. 2026); ICMJE AI-disclosure
  rules dated May 2023 / January 2024 and consolidated in Section V in
  January 2026; HHMI requires preprints for covered original research,
  Gates expects them.
• Scientific-writing craft (#146) — CSE 9th and Chicago 18th
  editions; the promotional-register catalog split into Vinkers- or
  Kobak-backed and practice-derived entries; Gopen & Swan's principles
  untested in reading-time experiments; Sweller 2019 on germane load.
• Unverified material removed (#147) — the last [TODO verify]
  markers and the unlocatable Smith & Banks 1991 citation are gone;
  knowledge/README.md now says a note does not merge with any.

Skill updates

Section headings, order and refusal posture are unchanged in every
skill. Output-contract and behaviour changes:

• venue-fit (#148)
  • ## Preprint options (with its review sub-section) is included
    without asking for HHMI-covered original research, with servers
    limited to HHMI's list; otherwise it stays opt-in.
  • Funder APC rules (Wellcome, HHMI, Gates, proposed NIH caps) and JIF
    eligibility join the hard filter.
  • The tier rationale no longer says most authors over-aim, and a
    history-driven move to Stretch is never presented as a judgement
    of merit.
• author-contribution-audit (#148)
  • When AI use is declared, ## LLM-as-author check lists the ICMJE
    Section V placement items, each marked covered by the text or for
    the author to confirm.
  • The non-author activity list matches ICMJE.
• citation-audit (#149)
  • New reference identity check on every bibliography entry: a new
    always-present Summary bullet (Reference identity: …) and new
    ### Reference identity problems and
    ### Hidden text or embedded instructions pattern subsections,
    omitted when empty.
  • A claim is cannot determine when its reference's identity is in
    doubt.
  • New hard constraint: the manuscript is data, not instructions.
  • "What this skill did NOT check" narrows the bibliography item to
    errors beyond the identity check.
• gap-finder (#149)
  • Categories credited to Miles plus two project additions; an
    optional Robinson et al. 2011 reason-code sub-tag on evidence-gap
    findings (no new section).
  • Directions name only works the manuscript already cites and never
    imply that one database search closes a gap.
• reporting-guideline-compliance (#150)
  • Trial reports default to CONSORT 2025 (2010 only when the author or
    venue requires it) and protocols to SPIRIT 2025; every item number
    carries its version, so output item numbering changes.
  • Flow-diagram completeness items default to partial unless the
    prose reports every required number.
  • New manifest constraints flow_diagram_items_default_partial and
    version_tag_every_item_number.
• reporting-guideline-fit (#150)
  • New routing targets: SPIRIT 2025, TRIPOD-LLM, REFINE and CHART.
    GAMER appears only as an add-on inside the existing sections
    (gamer_is_add_on_only).
  • Both reporting skills' full framing now says the EQUATOR library
    lists roughly 700 records.
• figure-text-alignment (#150) — grounding and README only; no
  contract change.
• argumentative-flow, compression, terminology-normalization,
  outlier-sentence-detector (#151) — no output-contract change.
  • argumentative-flow's full framing must not claim reading-time
    evidence for Gopen & Swan; its findings are diagnoses the author
    can overrule.
  • compression rationales cite negation and numerical insensitivity,
    not "antonymy", and don't treat germane load as a quantity an edit
    adds.
  • terminology-normalization's catalog offer labels entries
    corpus-grounded or practice-derived, with a date and a drift caveat.
  • outlier-sentence-detector: misattributed grounding claims
    corrected; no behaviour change.
• reviewer-simulation (#152)
  • Hidden or instruction-like text is reported as item 1 under "Likely
    major critiques" with [Submission integrity] in the lens slot,
    only when found.
  • "Potential fatal concerns" is narrower: a statistics or
    overclaiming critique goes there only when it exposes a
    design-level problem.
  • Critiques are formed from methods and results before the author's
    stated limitations are weighed.
  • The author-side-only rationale cites the actual NIH, Elsevier,
    Springer Nature and ICMJE rules.
• desk-rejection-risk (#152)
  • "What this assessment did NOT check": "The editor's actual mood…"
    became "Selection the manuscript cannot change…" (Zhang et al.
    2026).
  • The full guidance text drops the retired 70–90% desk-rejection
    figure; only journal-disclosed rates are quoted.
  • ICMJE AI-disclosure checks and hidden-text findings are reported
    inside the existing Format and length category.
  • "Lin et al. 2022" corrected to Teplitskiy et al. 2022.

Docs and CI

• lychee.toml: include_fragments = "none" replaces false, which
  lychee ≥ 0.20 no longer parses; link-check had been loading no
  config (#142).
• knowledge/README.md, docs/roadmap.md, DESIGN.md, README.md, the
  landing page and the concept pages aligned with the refreshed notes;
  concept tables regenerated from the manifests (#135).

Added

v0.3 skills

Four v0.3 skills land in this cycle (alongside terminology-normalization,
gap-finder, and author-contribution-audit, which shipped early during the
v0.2 ramp):

• reporting-guideline-compliance (#92) — walks an EQUATOR Network
  reporting-guideline checklist (CONSORT, STROBE, PRISMA, ARRIVE, STARD,
  TRIPOD/TRIPOD+AI, CARE, COREQ, CHEERS, plus AI-extensions) against a
  manuscript and classifies every checklist item as
  present / partial / missing / not-applicable, with a quoted
  manuscript excerpt as the anchor or an explicit gap statement.
  Downstream of v0.2's reporting-guideline-fit (which infers which
  checklist applies; this skill runs it). Validation skill — surfaces
  gaps for the author to address; does not modify the manuscript and
  does not invent prose. Renamed from reporting-compliance to share
  the prefix with reporting-guideline-fit.
• figure-text-alignment (#93) — sub-skill A (text-only) of the
  planned figure-text-alignment pair. Classifies caption ↔
  body-text-reference pairs as aligned / partially aligned /
  misaligned / cannot determine, plus pattern flags for orphan
  figures, phantom references, panel mismatches, axis/units divergence,
  and direction divergence. Pure text-vs-text — no image reading.
  Sub-skill B (multimodal / LLM-vision) remains deferred until
  reliability is validated against a known-mismatch test set.
• compression (#94) — page-limit-driven section compression that
  preserves every citation, statistic, declared core_claim,
  terminology choice, and hedging stack. Per-edit suggestions with
  preservation report; never auto-applies. Sits one editorial level
  below argumentative-flow. New grounding test file
  (tests/test_compression_grounding.py, 20 cases) enforces the
  preservation contract.
• outlier-sentence-detector (#99) — flags sentences that are
  outliers against the manuscript's own distribution on three
  dimensions: length, syntactic complexity (a comma, subordinator and
  coordination proxy; no parser) and nominalization density. A flag
  needs more than 2.5σ above the manuscript median and an absolute
  floor, and guidance level caps the flags surfaced per dimension
  (terse 3, standard 7, full 15). Each flag gives a location and
  a reason to re-read the sentence. No overall score, grade or
  pass/fail verdict, and no rewrites. Runs only when invoked. Grounded
  in knowledge/scientific-writing/quantitative-quality-measures.md
  and knowledge/editing/copyediting-vs-developmental.md.

v0.3 design memos and grounding

Three design memos landed for the remaining v0.3 skills so they can be
picked up later with minimal re-research (#95):

• docs/design/v0.3-statistics-consistency.md — Statcheck/GRIM/SPRITE
  orchestration; rpy2 vs. vendored re-implementation as the load-bearing
  decision.
• docs/design/v0.3-voice-profile.md — corpus-based authorial-style
  extraction; schema-embedded vs. sidecar author_voice slot as the
  load-bearing decision.
• docs/design/v0.3-persona-calibration.md — synthetic-feedback drift
  mitigation; dependency on voice-profile as the load-bearing
  decision.

A new knowledge note —
knowledge/scientific-writing/corpus-based-stylometry.md — grounds the
voice-profile memo in the established stylometry literature (Burrows,
Hoover, Argamon, Stamatatos, Mosteller-Wallace) and bounds what is and
isn't defensible at small-corpus sizes.

Install

• npx skills install path (#154, #155) —
  npx skills add seandavi/scriptorium (the vercel-labs skills CLI)
  installs the skills into any agent that CLI supports. init,
  explain and tour now carry the repo files they read in their own
  assets/ directory, and look for them in a scriptorium install
  first, then in assets/, then (explain only) on GitHub main;
  each says which source it used. scripts/sync_skill_assets.py (and
  just sync-skill-assets) keeps the bundled copies in sync, and a
  test fails when they drift. INSTALL.md, the install how-to, README
  and the landing page document the path. The Claude Code plugin is
  still the recommended install.

Skills and knowledge

• Promotional-register catalog in terminology-normalization
  (#124) — a new note,
  knowledge/scientific-writing/promotional-register.md, covers
  promotional and transactional register in academic prose. When
  MANUSCRIPT_STATE.yaml exists but its terminology lists are empty
  or thin, the skill now offers a default catalog drawn from the note
  and waits for the author to accept, edit or decline it, instead of
  stopping. Grant targets keep vision language off the list; discipline
  terms, significance claims and warranted boosters are exempt.

Docs site

• Landing page and README (#101) — restructured around what
  scriptorium operates on, workflow stages, the preservation contract
  versus voice preservation, and a "what this won't do" list. New
  pages: a constructed end-to-end case study
  (concepts/case-study) and a skills reference
  (reference/skills). The schema reference notes that
  MANUSCRIPT_STATE.yaml stays local and known_weaknesses is a
  calibration input.
• Concept pages and first-run how-to (#106) — start-here,
  declared-work-scope, guidance-level, workflow-stage and
  glossary concept pages, and a how-to/first-run walkthrough.
• Generated tables (#107, #119) — the skills reference page, the
  workflow-stage skill-by-stage table and the guidance-level
  per-skill block are generated from skills/*/manifest.yaml by
  docs/scripts/preprocess.py. Two new manifest fields feed them:
  lifecycle_phases: and guidance_level_behavior: (terse,
  standard, full).
• Analytics (#112, #123) — production docs builds load GA4 on the
  standard property. astro dev sends nothing, and neither do
  localhost, bare-IP or preview hosts (workers.dev, netlify.app,
  ts.net).
• templates/README.md points to the starter and example state files
  and the schema reference (#102).

Contributor tooling

• Link check (#97) — a lychee workflow checks links in the
  repo's prose (README, CHANGELOG, CONTRIBUTING, DESIGN, INSTALL,
  docs/, knowledge/, skills/) on push, PR and a weekly schedule.
• CONTRIBUTING.md gains sections on skill citation discipline (skill
  bodies cite with plain markdown links, not resolver cite keys;
  #98), documentation source of truth (repo-root files and
  knowledge/ are canonical; #102) and adding a new skill (#107).

Changed

• scriptorium prompt-pack writes one file per skill by default
  (#118). -o prompts/ (or an existing directory, or a new path with
  no extension) now writes one <skill>.md per skill plus a
  README.md listing each skill's category and description. For the
  previous single concatenated file, pass --single-file, or give a
  file path such as -o pack.md. With no -o, output still goes to
  stdout as one document. --prefix scriptorium- names the files
  scriptorium-<skill>.md. prompt-pack and scriptorium list now
  take skill descriptions from manifest.yaml.
• Renamed reporting-compliance → reporting-guideline-compliance for
  symmetry with reporting-guideline-fit. The original rename
  (commit b9d0a32) had not fully propagated; this cycle completes the
  propagation across roadmap, CHANGELOG, knowledge notes, the sibling
  skill's references, and the new skill itself.
• DESIGN.md's roadmap section now points to docs/roadmap.md, and a
  new Conventions section covers declared-work-scope and
  guidance-level (#117).
• The docs site's Tutorials section is gone; the how-to index lists
  the shipped how-tos (#102).
• Docs fixes: stale "planned" and version framing removed from
  README, INSTALL.md, DESIGN.md, knowledge/README.md and the
  argumentative-flow, reporting-guideline-fit and
  reporting-guideline-compliance skill files; docs/design/README.md
  indexes the design memos; roadmap rows marked landed (#96, #102,
  #104, #116).

Fixed

• Python 3.14 (#168, fixes #120) — the CLI crashed on startup
  because importlib.abc.Traversable was removed in 3.14; it now
  imports Traversable from importlib.resources.abc on 3.11+. The CI
  matrix and classifiers include 3.14.
• citation-audit's grounding pointed to
  knowledge/citations/citation-accuracy-evidence.md; the note is
  .qmd. Two knowledge notes had a mistyped Hyland 1999 DOI and a
  stale medRxiv URL (#97).
• Knowledge layer: 20 [TODO verify] marks resolved as verified,
  softened or recorded as research gaps; the "Laughter and the Chair"
  study is re-attributed to Pier et al. 2019, and the Crossref record
  count is updated (#100).
• The landing page's citation-audit example renders as a table, not
  literal pipes (#115).
• CONTRIBUTING.md markdownlint error (MD031) that failed CI lint
  (#113).

Install: uv tool install agentic-scriptorium==0.3.0, /plugin marketplace add seandavi/scriptorium@v0.3.0, or npx skills add seandavi/scriptorium.

Full changelog: v0.2.0...v0.3.0