Roadmap

The release plan for scriptorium, synthesized from the knowledge/ evidence base. Each phase’s contents are chosen because the research justifies them — not because they’re easy or because they sound impressive.

The implementation-priority section of every knowledge document feeds this roadmap. Findings that the research concluded should not become skills are documented in Explicit non-goals so the project’s claims stay honest.

Grounding lists below match the grounding: front matter in each shipped skill’s SKILL.md. Every skill also grounds in the two project-wide conventions, declared-work-scope and guidance-level; those are not repeated per row.

v0.1 — Foundation (released as 0.1.0.dev0, 2026-05-19)

The first release proves the architectural pattern. Three leaf skills, shared state, a CLI, and the knowledge layer that grounds the skills in evidence.

Component	Grounded in
citation-audit skill	citation-claim-alignment, citation-accuracy-evidence, citation-overreach-research, hallucination-in-llm-citations, reference-managers, peer-review-platforms
reviewer-simulation skill	reviewer-archetypes-evidence, common-critiques-taxonomy, ai-peer-review-research, critique-quality-evidence, reporting-guidelines, ai-writing-failure-modes, peer-review-platforms, credit-taxonomy-authorship
argumentative-flow skill	reader-expectation-approach, narrative-frameworks, argument-mapping, semantic-preservation, revision-research, esl-writers-swales-hyland
MANUSCRIPT_STATE.yaml schema + Venice example	All of the above
scriptorium CLI (install, validate, prompt-pack, list)	Self-evident
Claude Code plugin packaging	—
Knowledge layer (~40 docs)	—
DESIGN.md with scope statement + defensive-design section	ai-writing-failure-modes, discipline-conventions

Success criterion for v0.1: the three skills run usefully against the Venice 2026 manuscript (and any other manuscript with a populated MANUSCRIPT_STATE.yaml). Output structure is consistent enough that a future orchestrator can consume it.

v0.2 — Coordination + targeted critique additions (released as 0.2.0, 2026-05-20)

After v0.1 has been used on real manuscripts and the structured-output discipline is proven, the next priorities are coordination and the two highest-ROI critique additions identified in research.

Component	Status / Justification
manuscript-pipeline orchestrator	Not built; not in 0.2.0. Sequences leaf skills; consumes structured outputs. Spec ready; built once the leaves are stable. Should allow re-entry rather than a strict linear sequence (revision-research).
desk-rejection-risk skill	Landed. Most submissions to top general journals are declined without full peer review; the skill quotes only rates the journals disclose themselves and states that desk outcomes carry selection the manuscript cannot change. Grounded in editorial-decision-making, significance-positioning, common-critiques-taxonomy, peer-review-platforms.
venue-fit skill	Landed. Tiered venue recommendation with predatory refusal, funder APC rules in the hard filter, a preprint mode (PCI, Review Commons, F1000Research, eLife post-2022) that is opt-in except for HHMI-covered original research, and bias-managed pub-history calibration. Grounded in venue-selection, predatory-publishing, preprint-landscape, editorial-decision-making, significance-positioning.
ESL-aware checks embedded in argumentative-flow	Landed. esl-writers-swales-hyland.
author-contribution-audit skill	Landed. Replaces the originally-planned contributors: schema addition. Per declared-work-scope, scriptorium operates on declared prose where it lives — duplicating contributions in MANUSCRIPT_STATE.yaml would have created a sync problem. The skill audits the Author Contributions section against ICMJE’s four authorship criteria and CRediT’s 14 contributor roles, and checks AI-use disclosure placement against ICMJE Section V. Grounded in credit-taxonomy-authorship.
reporting-guideline-fit skill	Landed. Replaces the originally-planned reporting_guidelines: schema addition. Authors often don’t know which EQUATOR checklist applies — declaring it in state was the wrong-data-confidently-declared failure mode. The skill infers from the manuscript methods; the author confirms. Grounded in reporting-guidelines.

v0.3 — Validation skills + reporting-guideline compliance (released as 0.3.0, 2026-10-20)

0.3.0 ships figure-text-alignment (text-only), reporting-guideline-compliance, compression and outlier-sentence-detector; terminology-normalization and gap-finder shipped early, in 0.2.0. See [0.3.0] in the CHANGELOG. statistics-consistency, voice-profile and persona-calibration are deferred and not in 0.3.0; each has a design memo.

Once the structured-output pattern handles critique and transformation reliably, validation skills become the next leverage point. Most need deterministic scripts called out from skills, not LLM arithmetic.

Component	Status / Justification
statistics-consistency skill	Deferred; not in 0.3.0. statistical-inconsistency — Statcheck, GRIM, GRIMMER and DEBIT on extracted values; SPRITE is heuristic and not a verdict, and Carlisle-style corpus analysis is out of scope. LLM extraction feeds deterministic recomputation, and the report states coverage (results found vs. checked); the skill does not recompute in-band. Design memo: docs/design/v0.3-statistics-consistency.md.
figure-text-alignment skill	Shipped in 0.3.0 (text-only sub-skill A). visualization-figures, internal-consistency, forensic-methodology. Classifies caption ↔ body-text-reference pairs as aligned / partially aligned / misaligned / cannot-determine, plus pattern flags (orphan figure, phantom reference, panel mismatch, axis/units divergence, direction divergence). Text-vs-text by default. Reading figure images is opt-in through the figure-image-reading scope guard, and every image-derived finding is then labelled “unvalidated vision”. A validated multimodal sub-skill B (LLM-vision) remains deferred until reliability is validated against a known-mismatch test set; current multimodal benchmarks (PRISMM-Bench, MuSciClaims) support the deferral. Image-integrity claims stay out of scope (the no-image-integrity-claims accuracy guard, not overridable).
terminology-normalization skill	Shipped early, in 0.2.0. internal-consistency, style-guides, promotional-register, significance-positioning — terminology drift detection; preferred-term enforcement from MANUSCRIPT_STATE.yaml; an opt-in promotional-register catalog whose entries are labelled corpus-grounded or practice-derived. Shipped during v0.2 ramp because the grounding notes existed and the schema fields (terminology.preferred / forbidden / synonyms) were already in place.
gap-finder skill	Shipped early, in 0.2.0. Identifies gaps in declared draft prose, organised by seven categories: five adapted from Miles (2017; Adu & Miles 2023) plus two manuscript-internal additions (counterargument, internal-consistency). Each finding anchors in a specific manuscript passage; suggested directions are pasteable search strategies by default, never invented citations. Named papers are opt-in through the citation-recommendations scope guard and limited to works resolved against Crossref or PubMed in the session. Grounded in research-gap-detection, literature-search-strategies, argument-mapping, logical-fallacy-detection, internal-consistency, common-critiques-taxonomy, hallucination-in-llm-citations.
reporting-guideline-compliance skill	Shipped in 0.3.0. reporting-guidelines, internal-consistency. Walks an EQUATOR Network checklist (CONSORT, SPIRIT, STROBE, PRISMA, ARRIVE, STARD, TRIPOD/TRIPOD+AI, CARE, COREQ, CHEERS, plus AI-extensions) item by item and classifies each as present / partial / missing / not-applicable, with quoted manuscript anchors. Trial reports default to CONSORT 2025 and protocols to SPIRIT 2025; every item number carries its checklist version. Downstream of v0.2’s reporting-guideline-fit (which infers which checklist applies; this skill runs it).
compression skill	Shipped in 0.3.0. narrative-frameworks, semantic-preservation, copyediting-vs-developmental, revision-research, reader-expectation-approach, hayes-flower-writing-model, esl-writers-swales-hyland, ai-writing-failure-modes. Page-limit-driven section compression that preserves every citation, statistic, declared core_claim, terminology choice, and hedging stack. Per-edit suggestions; never auto-applies. Sits one editorial level below argumentative-flow (line-edit posture vs. block-rewrite).
outlier-sentence-detector skill	Shipped in 0.3.0. quantitative-quality-measures, copyediting-vs-developmental. Flags sentences that are outliers against the manuscript’s own distribution on length, syntactic complexity and nominalization density, for human review. No quality score; does not edit.
voice-profile skill	Deferred; not in 0.3.0. corpus-based-stylometry, author-role-evidence — extract author writing patterns from a small single-author corpus. Design memo: docs/design/v0.3-voice-profile.md.
persona-calibration skill	Deferred; not in 0.3.0. author-role-evidence, ai-peer-review-research — checkpoint synthetic feedback against the real author. Design memo: docs/design/v0.3-persona-calibration.md.

v0.4 — Grant-specific skills and bounded transformations

Originally framed as “Generation skills”. Reframed once declared-work-scope landed: by default, scriptorium does not generate prose from blankness. Generation is in scope when it transforms declared scaffolding into a known target form — the v0.4 work concentrates there. Most of the originally-planned v0.4 skills (discussion-drafting, results-narrative) involved substantial proposer-side judgment incompatible with the scope and have been dropped (see Explicit non-goals); two transformations remain in scope and one grant-specific critique is added.

The unifying frame for v0.4 is the grant-writing workflow, where the author has typically done substantial proposer-side work (mentor discussions, aim selection, significance framing) before sitting down to write, and where the value of bounded-transformation skills is highest.

Component	Justification
specific-aims skill	Transforms declared significance + hypotheses + methods (in MANUSCRIPT_STATE.yaml and the manuscript’s existing methods scaffolding) into structured NIH-aims prose. Grant-specific; the author has typically already committed to the aims via mentor discussion before invoking — the skill renders them in canonical structure. Grounded in significance-positioning and nih-significance-patterns. NIH 2025 Simplified Review Framework bundles Significance + Innovation — the skill ladders both, and the NIH Factor 1 / Factor 2 framing is the target structure. In scope because the inputs are declared. For NIH targets, rendering application prose sits behind the nih-application-prose policy guard (default: ask, source NIH NOT-OD-25-132); see the decision below.
aims-significance-coherence skill	Critique skill — audits whether the declared significance is coherent with the stated aims (cross-section consistency check for grants). Pairs with reviewer-simulation’s grant-archetype variant (study-section roles); the aims-significance gap is one of the most common NIH-reviewer flags. Grounded in nih-significance-patterns, significance-positioning, and reviewer-archetypes-grants.
lay-summary skill	Translation of declared manuscript or grant prose into plain-language form, against funder-specific length and reading-level requirements (NIH 2025 Public Access requirements; Wellcome Trust; EU Clinical Trials Regulation 536/2014). Strongly transformative — both source and target style are declared. Grounded in plain-language-lay-summaries.

Inputs from the October 2026 knowledge refresh

The refresh (#125) changed the grounding for the unshipped v0.4 work. These are inputs to the v0.4 design, not shipped behaviour:

• specific-aims (nih-significance-patterns, #138): NIH NOT-OD-25-132 says NIH will not consider applications, or sections of applications, that are substantially developed by AI to be the applicants’ original ideas. The note recommends that, for NIH targets, the skill default to a slot audit plus author-written suggestions, with prose rendering only on explicit request and labelled with the policy.
• aims-significance-coherence (nih-significance-patterns): check SRF Factor 1 (importance) first, because NIH’s scoring guidance makes it a ceiling on Overall Impact. As pure critique, it cannot conflict with NOT-OD-25-132, which makes it the safer first v0.4 deliverable.
• Grant reviewer personas for reviewer-simulation (reviewer-archetypes-grants): use the funder’s current discussion line (the top third under NIH’s emergency rules); check NSF Broader Impacts against the 2025 priorities guidance; date the persona text; never map scores to paylines or funding probabilities (that stays fixed, as the no-numeric-outcome-probability accuracy guard). Simulating or drafting reviews of applications the user was assigned to review (NOT-OD-23-149) becomes a policy guard rather than a fixed refusal: the skill states the concern and its source and the author decides. The shipped editorial-side-use guard (default: ask) already covers reviewers and editors of submitted work and cites NOT-OD-23-149.
• terminology-normalization venue-manual layer (style-guides, #146): a possible opt-in layer beneath the author’s declared terminology lists, citing the manual and edition for each suggestion. Maybe later; it needs its own design decision.

Decision (2026-10-20, #156): specific-aims may render NIH application prose, behind the nih-application-prose policy guard (default: ask, source NIH NOT-OD-25-132). For an NIH target the skill states the NOT-OD-25-132 concern in one line and asks before rendering; an author can record a standing answer under meta.overrides. This follows the author-authority convention: whether a given use is “substantially developed by AI” is a judgement the author makes and answers for. The guard id is reserved for specific-aims; scriptorium validate rejects overrides of it until the skill ships.

Skills outside this list that were on the previous v0.4 roadmap — results-narrative, discussion-drafting — have been dropped under declared-work-scope. The remaining grant-side work (cover letters, biosketch-fit, funder-specific compliance checks) is plausibly v0.5+ once v0.4 lands; not committed yet.

v0.5+ — Platform reach + knowledge expansion

Component	Justification
Codex / Gemini / Hermes adapters	Audience reach beyond Claude Code. Most reusable via the prompt.md files already shipped per skill; thin per-platform installer scripts in adapters/.
Per-discipline knowledge layers (physics, CS/ML, mathematics, qualitative social science)	discipline-conventions — currently scope-limited to biomedical/clinical. Expand only when non-biomedical adoption emerges.
Astro/Starlight docs site (with Quarto preprocessing)	Landed. Built from docs/ (docs/scripts/preprocess.py copies the knowledge layer and renders Quarto) and deployed to GitHub Pages by docs-deploy.yml.

Explicit non-goals

Findings the research concluded should not become skills, with reasons. This list is load-bearing: it keeps the project honest about what it does and doesn’t claim.

• No blank-slate prose generation by default — declared-work-scope. Scriptorium operates on prose the author has written or scaffolding the author has declared. The proposer role in Hayes’ 2012 writing-process model (generating content from nothing) is the author’s; scriptorium occupies the translator and evaluator roles. Generation skills are in scope when they transform declared scaffolding (v0.4 specific-aims, lay-summary). Generation from a blank section, or “help me figure out what to write about”, is not, and the originally-planned v0.4 discussion-drafting and results-narrative skills were dropped under this scope: the discussion involves substantial proposer judgment (what does this mean? what should be emphasized?) and the “results narrative” risks slipping claims that go beyond the declared data. The shape of those concerns is covered by argumentative-flow + gap-finder (for an existing discussion) and figure-text-alignment (for results-prose-vs-data consistency). The boundary refusals in shipped skills are scope guards with default: refuse; under author-authority an author can override them, but no skill is built for blank-slate drafting.
• No general-purpose writing-quality score — quantitative-quality-measures. Flesch-Kincaid / SMOG / Coleman-Liau systematically misrate scientific prose (technical terms inflate difficulty). A quality score would be theater.
• No authorial-voice preservation guarantee — ai-writing-failure-modes. Detection of “ChatGPT smell” (Kobak et al. 2025) is possible at corpus level; correction at sentence level is unreliable. Scriptorium’s conservative-edit posture mitigates this but doesn’t claim to eliminate it.
• No forensic-expert replacement — forensic-methodology. Bik-style image forensics, Cabanac tortured-phrase detection, and statistical forensics (Carlisle, Statcheck, GRIM, SPRITE) require domain experts. Scriptorium is a pre-submission first pass that catches cheap errors before a manuscript reaches reviewers — not a replacement for sleuths or institutional integrity review.
• No autonomous reviewing — ai-peer-review-research, peer-review-platforms. Scriptorium’s reviewer-simulation and desk-rejection-risk are explicitly author-side: the author runs them on their own work to pressure-test before submission. Editorial-side use would mean uploading a confidential manuscript and delegating the critique to AI, which current reviewer policies restrict. Run by a reviewer or editor of a submitted manuscript, those skills (and venue-fit and author-contribution-audit) apply the editorial-side-use policy guard (default: ask): they state the concern and its sources and ask before proceeding. NIH prohibits its reviewers from using generative AI to analyse or formulate critiques of grant applications (NOT-OD-23-149). Elsevier and Springer Nature (2026) allow limited, disclosed, supportive reviewer use but forbid delegating the critique to AI. ICMJE requires reviewers to follow the journal’s AI policy or request the journal’s permission. We agree with those policies.
• No replacement for reference managers — reference-managers. Citation auditing works with whatever bibliography Zotero/Mendeley/Paperpile/BibTeX produces; scriptorium does not manage references itself.
• No discipline-specific defaults beyond biomedical/clinical at v0.1–v0.3 — discipline-conventions. The evidence base is biomedical-coded. Expanding to physics, CS/ML, math, humanities requires per-discipline knowledge layers that don’t exist yet; PRs welcome.

Update cadence

This roadmap is reviewed at each release. Issues open against deferred items are welcome but get triaged against the priority order above.