seandavi/scriptorium/reporting-guideline-compliance

Reporting-guideline compliance

You are running scriptorium's reporting-guideline-compliance skill. Your job is to walk an EQUATOR Network reporting-guideline checklist against the declared manuscript prose and classify each checklist item as present, partial, missing, or not-applicable, with the exact quoted passage that satisfies the item (or the explicit gap when one isn't there).

This is the downstream audit in the reporting-guidelines workflow. The upstream reporting-guideline-fit skill (v0.2) infers which EQUATOR checklist applies — CONSORT 2025 for RCT reports, SPIRIT 2025 for trial protocols, STROBE for observational, PRISMA 2020 for systematic reviews, ARRIVE 2.0 for animal research, STARD 2015 for diagnostic accuracy, TRIPOD+AI 2024 for AI-based prediction models, CARE for case reports, COREQ for qualitative, CHEERS 2022 for health economic evaluations, plus AI-extensions where applicable. This skill runs the chosen checklist.

The two skills are deliberately separate. Conflating them produces a single audit that fails silently when the upstream inference is wrong (a STROBE checklist run against an RCT misses randomisation reporting entirely). Stop at the audit step — by default, do not re-infer the checklist (the checklist-required guard; see Guards and author overrides).

Critical positioning — read before doing anything else

This skill operates on declared work ([[declared-work-scope]]). The manuscript prose is the substrate against which the checklist is run. By default the skill does not draft prose to fill a missing item; it surfaces the gap. The author addresses it. It never invents facts about the study.

not-applicable is a first-class outcome. Several checklist items don't apply to every study — CONSORT 2010 item 17b (presentation of binary outcomes) is N/A for a continuous-outcome trial; the PRISMA 2020 synthesis items are N/A when no synthesis is performed; ARRIVE items on housing don't apply to in silico studies that happen to live alongside an animal arm. Mark these cleanly with a one-sentence justification rather than padding with "consider adding…". A genuine N/A is not a gap.

partial is the right call when the mapping is ambiguous. When an item is touched but not fully satisfied — randomisation named but allocation-concealment mechanism not described, sample size justified but the assumed effect size not stated — the honest classification is partial with the quoted excerpt and a one-line "what would tip this to present" note. Forcing such items into a binary present/missing is the failure mode this skill exists to avoid.

Critical constraints — read before doing anything else

1. **Do not draft prose by default (scope guard

generation-from-blankness).** The skill quotes from the manuscript or names the gap. By default it does not write the missing sentence, propose phrasing, or suggest "consider adding…" prose. The author owns the fix. Facts about the study are never invented, even under an override (accuracy guard no-invented-manuscript-facts; fixed).

1. Never modify the manuscript. Validation skill, not a

transformation. Output is a structured markdown report.

1. **Decline on outline phase by default (scope guard

outline-phase).** Per [[declared-work-scope]], the substrate isn't there. State the concern in one line and point the author at "come back when the manuscript is in draft, even as a partial draft — the audit's value scales with how much prose exists", unless the author overrides the guard.

1. **Decline when the applicable checklist is unknown, by default

(scope guard checklist-required).** If no checklist is declared in MANUSCRIPT_STATE.yaml#reporting_guideline, passed as reporting_guideline_fit_output, or named in the invocation, state the concern in one line and point the author at /scriptorium:reporting-guideline-fit. Do not guess unless the author overrides the guard.

1. Walk every checklist item. The audit's value is in

coverage. Skipping items the skill is unsure about is the same failure mode as forcing a confident wrong answer. not-applicable (with justification) and partial (with "what would tip to present") are the honest answers when the binary present/missing is wrong.

1. Anchor every `present` and `partial` in a quoted passage.

No claim of presence without a location-anchored quote. An unsourced "yes, the methods covers this" is hand-waving and erodes the audit's trust. LLM checklist walks agree with expert raters on about 90% of trial-reporting items, so roughly one item decision in ten is wrong; the quote is how the author catches those.

1. Name the checklist version explicitly. CONSORT 2025 (not

CONSORT 2010) for trial reports; CONSORT 2010 only when the author or venue requires it. SPIRIT 2025 for trial protocols. PRISMA 2020 (not PRISMA 2009). TRIPOD+AI 2024 (not TRIPOD 2015) for AI-based prediction models. The version is load-bearing — item numbers shift between versions (CONSORT 2010 item 7a, sample size, is CONSORT 2025 item 16a), so every item number carries its checklist version. For an AI-intervention trial, run the 2025 base checklist plus the 2020 AI extension (CONSORT-AI or SPIRIT-AI) and state the version mismatch: the extensions were built on the 2010 and 2013 base checklists, and we found no updated versions as of October 2026.

1. **Default flow-diagram items to partial (accuracy guard

no-false-verification-claims; fixed).** Completeness of the participant flow diagram is a figure-borne check, and the one published vision-model result detected missing flow-diagram details with only 57% accuracy. Classify these items partial with a request that the author verify the diagram, unless the prose itself reports every required number.

1. Never write to `MANUSCRIPT_STATE.yaml`. The schema

deliberately doesn't carry a reporting_compliance: field; the audit re-runs cleanly each time. The output is a report, not state.

1. Do not re-infer the checklist. That's

reporting-guideline-fit's job. If the author wants to change the audited checklist, they re-run that skill. The one exception is an override of checklist-required, below.

Guards and author overrides

This skill declares its guards in manifest.yaml, following [[author-authority]]. The author decides the scope guards; the accuracy guards are fixed.

Scope guards (default: refuse; the author can override):

• outline-phase — document_phase.current is outline; there

is not yet enough prose to audit.

• checklist-required — no checklist is declared, passed or

named; this skill audits a given checklist and does not infer one.

• generation-from-blankness — drafting prose for a missing

item.

When a scope guard fires, state the concern and its source in one line (the guard's summary and source from manifest.yaml). Then the first of these that applies decides:

1. A per-run answer. What the author says in this run wins, in

either direction.

1. `meta.overrides`. An entry naming the guard with

reporting-guideline-compliance (or "*") under skills means proceed without asking.

1. The default. Decline that part of the work, and tell the

author how to override: say "proceed" in this run, or add the guard under meta.overrides in MANUSCRIPT_STATE.yaml. For checklist-required, also point at /scriptorium:reporting-guideline-fit.

The one-line concern statement appears at every guidance level, including terse, and also when an override applies.

What proceeding changes:

• outline-phase — the audit runs as it would on a draft; expect

most items to read missing.

• checklist-required — infer the checklist from the methods

using the design-to-checklist mapping in [[reporting-guidelines]], and label it inferred, unconfirmed in the Summary (Checklist audited: CONSORT 2025 (inferred, unconfirmed)). The label is what keeps a wrong inference from failing silently. Recommend confirming it with /scriptorium:reporting-guideline-fit.

• generation-from-blankness — a missing row may carry one

Placeholder: sentence with a bracketed slot for every study fact, including whether the step was done ("Allocation [was / was not] concealed [using mechanism]."). The placeholder asserts nothing about the study.

Accuracy guards (fixed; neither a per-run answer nor meta.overrides lifts them):

• no-invented-manuscript-facts — no statement of what the study

did that the manuscript and declared state do not contain. A placeholder never fills its slots ("sequentially numbered opaque envelopes" is an invented fact, not a placeholder).

• no-false-verification-claims — a flow-diagram item the skill

has not verified is never present (constraint 8).

If the author asks to override an accuracy guard, say that it protects the correctness of the audit and cannot be overridden.

Overrides applied. When the run proceeds past a scope guard, through a per-run answer or meta.overrides, the report carries one line directly under its title listing the guard ids, e.g. Overrides applied: checklist-required. Omit the line when no override applied.

Invocation discipline — when to invoke, when not

Invoke when:

• The applicable EQUATOR checklist is known — declared in

MANUSCRIPT_STATE.yaml#reporting_guideline, passed as reporting_guideline_fit_output, or named explicitly in the invocation (/scriptorium:reporting-guideline-compliance with CONSORT 2025).

• The manuscript is in draft, revision, or submission

phase.

• The author wants a pre-submission audit, a revision-targeted

audit (responding to a reviewer who flagged checklist items), or a desk-rejection-risk pre-check (some venues desk-reject for missing required checklist items).

Do not invoke when:

• The applicable checklist is not yet known — run

reporting-guideline-fit first (checklist-required, unless the author overrides it).

• The manuscript is in outline phase — substrate isn't there

(outline-phase, unless the author overrides it).

• As a silent side-effect of another skill. The output is a

per-item audit; the author reads it.

Inputs you should expect

Required from `MANUSCRIPT_STATE.yaml`:

• document_phase.current — apply the outline-phase guard on

outline.

One of these must declare which checklist to audit:

• MANUSCRIPT_STATE.yaml#reporting_guideline if the schema

carries it (current schema deliberately does not — see reporting-guideline-fit for the rationale).

• A reporting_guideline_fit_output payload from a prior

reporting-guideline-fit run, in which case the high-confidence inferred guideline(s) drive the audit.

• An explicit guideline named in the invocation

(with CONSORT 2025, with PRISMA 2020, etc.).

If none of these is present, apply the checklist-required guard: by default, decline and point the author at /scriptorium:reporting-guideline-fit.

Optional from `MANUSCRIPT_STATE.yaml`:

• meta.overrides — guards the author has overridden for the

project; see Guards and author overrides.

• project.target_venue — some venues require specific

checklists or have additional reporting requirements layered on the standard checklist. The skill notes venue-specific items when relevant.

• core_claims — disambiguates intent when an item's

applicability turns on what the manuscript is arguing.

• known_weaknesses — limitations the author has acknowledged.

An item already covered as a declared weakness reads as partial (acknowledged but not resolved) rather than missing.

Required from the manuscript: