Declared-work scope

import { Aside } from ‘@astrojs/starlight/components’;

Scriptorium operates on prose the author has already written, or on scaffolding the author has committed to in MANUSCRIPT_STATE.yaml. By default it does not produce prose from a blank page. This page is the why-and-where of that default, and the table at the bottom shows how each shipped skill respects it.

The boundary is a default the author controls. Each skill declares its boundary refusals as scope guards with default: refuse, and the author can override a scope guard for one run or for the project. The accuracy guards that sit alongside them, such as never citing from model memory, cannot be overridden. Author authority covers the mechanism.

The fuller, evidence-grounded version of this principle lives in knowledge/conventions/declared-work-scope. This page is the author-facing summary.

The scope cut

Scriptorium reads two kinds of input:

1. Declared prose. Sentences, paragraphs, sections you have already written — even rough or unfinished. A methods section in draft. A discussion paragraph that doesn’t yet flow. An abstract stub.
2. Declared scaffolding. Anything you have committed to in MANUSCRIPT_STATE.yaml — the core_claims you are arguing, the known_weaknesses you have already chosen to acknowledge, terminology.preferred / terminology.forbidden, the target_venue, the audience, the tone, the voice.

Both count as declared work. A skill grounds its output against one or both. It does not synthesise output from nothing.

The mirror default: scriptorium does not draft sections from a blank page. It does not propose what the paper should be arguing. It does not invent claims for the author to defend. A skill steps outside this only when the author overrides the scope guard that holds it.

A future skill that takes declared significance + declared hypotheses + declared methods and turns them into specific-aims prose would still be in scope, because the inputs are declared scaffolding. (For NIH applications it would also carry the `nih-application-prose` policy guard; see the [roadmap](/roadmap/).) A skill that helps an author "figure out what to study" is not in scope by default, because there is nothing declared to ground against. The principle is *declared inputs*, not *read-only*.

Why this cut

Two reasons, one cognitive and one empirical.

Cognitive: Hayes 2012

The Hayes-Flower cognitive-process model of writing, in its 2012 revision, splits the writing process into four sub-processes — proposer, translator, transcriber, and evaluator — arbitrated by a control structure that decides which one to engage next. The model is the canonical psychological account of what writers actually do.

Scriptorium’s scope maps cleanly onto Hayes’ processes:

• Proposer — generating content from nothing — is out of scope. This is the activity that requires intuition, domain knowledge, and the author’s own judgement.
• Translator — turning declared proposals into language — is in scope. argumentative-flow lives here.
• Transcriber — orthography — is trivially in scope.
• Evaluator — assessing what’s been written — is in scope. citation-audit, reviewer-simulation, gap-finder, and desk-rejection-risk are all evaluator-role skills.

Hayes’ updated model treats working memory as the writer’s bottleneck. The discipline scriptorium adopts is to absorb the translator and evaluator load so the author’s working memory stays available for the proposer role. Stepping into the proposer role would defeat the purpose.

For the full grounding see hayes-flower-writing-model in the knowledge layer, and the workflow-stage page for the practical implications.

Empirical: the failure modes of blank-slate generation

The two most-cited failure modes of LLM scholarly writing — hallucinated citations and authorial-voice loss — are both failure modes of blank-slate generation. They happen much less, or do not happen at all, when the model is operating on declared inputs.

• A skill that audits the author’s declared citations against bibliographic metadata cannot hallucinate references. A skill that generates prose from blankness will, eventually, hallucinate references to back unbacked claims.
• A skill that operates on the author’s existing prose has the author’s voice as input and can preserve it. A skill that generates from blankness produces centre-of-distribution prose by construction.

Scriptorium’s declared-work scope is the structural defence against these failure modes. The per-skill guards — “no invented citations” for citation-audit, “preserve hedging and stance markers” for argumentative-flow, “never claim verification of papers you haven’t seen” for any skill that touches references — are the same principle expressed at each skill’s surface.

The deeper grounding for the failure-mode evidence is ai-writing-failure-modes.

What “declared” means in practice

The bar for “declared” is much lower than the bar for “finished”:

• A draft sentence — even if you plan to rewrite it.
• A stated core_claim in MANUSCRIPT_STATE.yaml.
• A known_weakness listed in MANUSCRIPT_STATE.yaml, even before the limitations paragraph is written.
• A paragraph stub that names what it will argue but not yet how.
• A methods section that’s complete in structure but still rough in prose.
• A target_venue declared in project.target_venue, even before a draft exists for that venue’s format.

In all of these cases, scriptorium has something to ground against and can run. The trigger for the default refusal is not “this prose is rough,” it is “there is no prose here yet.”

What’s out of scope by default

What the skills decline unless the author overrides the scope guard involved:

• Draft an empty section. Asking a skill to fill in an empty Related Work section is a blank-slate generation request (generation-from-blankness). By default the skill declines and points at the pre-declaration work that needs to happen first.
• Write a discussion from a results table. The discussion is proposer work — it requires the author’s reading of what the results mean. A skill cannot ground discussion prose against a table alone.
• Propose new aims, claims, or hypotheses. Adding to core_claims based on what would “round out” the paper is proposer work the author owns (pre-declaration-ideation).
• Generate a literature review from a topic name. Even if a bibliography is supplied, choosing what to engage with — and how — is proposer work.

One related refusal is not a scope guard and no override changes it:

• Invent a citation to support a claim. This is the canonical hallucination case, and no-parametric-citations is an accuracy guard: no skill names a citation from model memory. By default citation-audit and gap-finder surface missing support and suggest search strategies, and the author chooses what to cite. An author who overrides their citation-recommendations scope guard gets named papers, but only works the skill resolved against Crossref or PubMed in that session.

When a skill is asked to operate beyond its declared-work boundary, the refusal pattern is to name the refusal clearly, point at the pre-declaration work, say how to override (answer in this run, or add an entry under meta.overrides), and not lecture. One sentence is enough.

The phase axis is a proxy

MANUSCRIPT_STATE.yaml has a document_phase.current field that takes values outline / draft / review / revision / submission / post-submission / accepted. Several skills decline by default on outline (the outline-phase scope guard) for the same reason this page describes: outline-phase manuscripts often have empty sections, and operating on an empty section is blank-slate generation.

The phase axis is a proxy for the declared-work principle. It catches the common cases. It misses edge cases — a draft-phase manuscript with an empty Related Work section is still asking for blank-slate generation; a revision-phase author asking about limitations framing is in scope even before the limitations paragraph is written. The phase axis is too coarse to be the actual rule; the declared-work axis is.

For the operational rule each skill follows on the phase axis, see the skills reference and the workflow-stage page.

How this shows up in skills

The table is the operational expression of this page. Every shipped skill names what it consumes from the author’s declarations and what it declines to synthesise by default. Rows are alphabetical. Which of these are overridable scope guards and which are fixed accuracy guards is listed per skill in the guard table.

Skill	Declared inputs	What it won’t invent by default
argumentative-flow	Manuscript section + terminology.* + style.tone / style.voice + declared core_claims	New claims; new citations; voice changes outside declared style
author-contribution-audit	Author Contributions section + project.target_venue (optional)	Authorship adjudication; CRediT roles not derivable from declared text
citation-audit	Manuscript prose + bibliography	New citations; full-text claims about papers not supplied
compression	Manuscript section + constraints.max_word_count + declared core_claims	Cuts to declared claims; cuts to citations or statistics
desk-rejection-risk	Manuscript + project.target_venue	Generic advice (declines without a declared venue)
explain	Plugin tree only	Anything outside the plugin tree
figure-text-alignment	Manuscript prose + figure captions	Image content (text-only by default); fabricated panel descriptions
gap-finder	Manuscript prose + declared core_claims + known_weaknesses	Prose to fill the gap; citations to fill it
init	Filesystem + author Q&A	Fields beyond what the schema declares
outlier-sentence-detector	Manuscript prose only	Universal quality thresholds; the calibration is per-manuscript
reporting-guideline-compliance	Manuscript + inferred reporting guideline	Prose to fill missing checklist items
reporting-guideline-fit	Manuscript methods + abstract	Authoritative declaration of the guideline; the author confirms
reviewer-simulation	Manuscript + core_claims + known_weaknesses + project.target_venue	Critiques of work not in the manuscript
terminology-normalization	Manuscript + terminology.* lists	Stylistic flattening beyond the declared lists
tour	Plugin tree only	Anything beyond what’s in the plugin tree
venue-fit	project.target_type + manuscript metadata + (optional) pub_history	Predatory venues; venue assessment without target_type

The pattern that repeats: the declared inputs column is what the skill grounds against; the won’t invent column is what the skill does not produce by default because it has no declared ground to anchor against.

Related

• Start here — the conceptual map this page is one branch of.
• Workflow stage — the practical expression of the proposer/translator/evaluator split.
• Guidance level — the sister convention controlling how scriptorium talks; this page controls where it operates.
• Author authority — how the author overrides this page’s scope guards, and which guards stay fixed.
• knowledge/conventions/declared-work-scope — the full, evidence-grounded note.
• hayes-flower-writing-model — the cognitive-process model this scope cut grounds in.
• ai-writing-failure-modes — the empirical evidence the scope cut defends against.