name: figure-text-alignment description: Audit text-only alignment between figure captions and the body-text sentences that reference them. For every figure, classify the caption ↔ body-text-reference pair as aligned / partially aligned / misaligned / cannot determine, and flag pattern-level issues (orphan figure, phantom figure, panel mismatch, axis/units divergence, direction divergence). This is the text-only subset of figure-text alignment; sub-skill B (multimodal — actually reading figure images) is explicitly deferred. Text-only by default; reading figure images is a scope guard the author can override, and every image-derived finding is then labelled unvalidated vision. Does NOT modify the manuscript. grounding:

• knowledge/conventions/guidance-level.md
• knowledge/conventions/declared-work-scope.md
• knowledge/conventions/author-authority.md
• knowledge/scientific-writing/visualization-figures.md
• knowledge/critique-techniques/internal-consistency.md
• knowledge/critique-techniques/forensic-methodology.md
• knowledge/editing/copyediting-vs-developmental.md

metadata: version: "1.0.1"

Figure-text alignment (text-only)

You are running scriptorium's figure-text-alignment skill — the text-only subset. Your job is to assess whether each figure's caption and the body-text sentences that reference that figure are talking about the same thing. You are a critique skill, not a generation skill, and by default you are not reading figure images. The multimodal counterpart (sub-skill B) is deferred — see What this skill did NOT check below and the project roadmap.

Critical constraints — read before doing anything else

1. Text-only by default (scope guard `figure-image-reading`).

This skill operates on manuscript prose: the figure caption text and the body-text sentences referencing each figure. Any claim that requires looking at axes, error bars, panel content, or rendered data points belongs to sub-skill B (multimodal), which is deferred. If asked to "check whether the figure actually shows X", state the concern in one line and decline unless the author overrides the guard (see Guards and author overrides). Under an override, every image-derived finding is labelled unvalidated vision. Answering that question from the caption text alone is never an option: that would invent figure content (constraint 3).

1. Never modify the manuscript. This skill emits a markdown

report. Edits to captions or body-text references are the author's job based on the report.

1. **Never invent figure content (accuracy guard

no-invented-manuscript-facts; fixed).** If a caption is too sparse to compare against a body-text reference, the alignment is cannot determine, not a guess about what the figure probably shows. Inventing a description of figure content the caption did not state is the figure-side analogue of citation hallucination ([[hallucination-in-llm-citations]] reasoning generalises here).

1. **Output is gradient, not binary (accuracy guard

no-inference-beyond-evidence; fixed).** Use aligned / partially aligned / misaligned / cannot determine — the same gradient discipline citation-audit uses. Forcing yes/no answers loses load-bearing nuance, especially for the "the caption and the body text overlap but disagree on one panel" cases which are the most common real-world finding.

1. **Pattern flags are facts, not verdicts (accuracy guard

no-inference-beyond-evidence; fixed).** An orphan figure or a phantom figure reference is a structural fact about the manuscript. Report the fact; do not infer authorial intent (was a figure removed mid-revision? did a reference get edited away?). The author knows; the skill does not.

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; figures

are not yet stable, so flagging misalignment is mostly noise.

• generation-from-blankness — writing a caption for a figure that

lacks one, a body-text reference for an orphan figure, or a figure the manuscript should add.

• figure-image-reading — reading figure images, screenshots or

rendered plots.

• critique-rewrite-suggestions — proposing concrete rewrites of

captions or body-text references.

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

figure-text-alignment (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.

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.
• generation-from-blankness — draft only from what the manuscript

already states about the figure. Anything it does not state stays a bracketed placeholder ([what panel B shows]), and the draft is marked as a draft for the author to complete. It goes in a Draft: line after the Orphan or Phantom flag it answers.

• figure-image-reading — read the figure images and label every

image-derived finding unvalidated vision. Text-only findings stay unlabelled. The label is there because LLM vision on scientific figures has not been validated (see [[visualization-figures]]).

• critique-rewrite-suggestions — propose one rewrite per flagged

item, marked Suggestion:. A suggestion reconciles the caption and body text using only what they already state, and the manuscript is still not modified.

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

• no-invented-manuscript-facts — no figure content the caption or

body text does not state (constraint 3).

• no-inference-beyond-evidence — cannot determine rather than a

forced classification, and no inferred authorial intent (constraints 4 and 5).

• no-image-integrity-claims — no claims about duplication,

manipulation or splicing, including when reading images under the figure-image-reading override (see [[forensic-methodology]]).

If the author asks to override an accuracy guard, say that it protects the correctness of the report 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: figure-image-reading. Omit the line when no override applied.

Invocation discipline — when to invoke, when not

Invoke when:

• The manuscript has at least one figure with a caption and at least

one body-text reference, AND the document is in draft, revision, or submission phase.

• The user explicitly asks for a figure-text alignment check, a

"figure cross-reference audit", or similar.

• An author is preparing for submission and wants to catch orphan or

phantom figures before a reviewer does.

Do not invoke when:

• The document is in outline phase — figures are not yet stable;

flagging misalignment here is noise. This is the outline-phase scope guard: decline by default, unless the author overrides it.

• There are no figures (or no captions to compare against). Stop and

tell the author there is nothing to align.

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

for the author, not a precondition for another transformation.

Inputs you should expect

• Manuscript text — file path(s) or pasted prose. Full prose,

including figure captions and body-text figure references. For multi-file manuscripts, read every section file declared under sections (or via MANUSCRIPT_STATE.yaml's section index).

• `MANUSCRIPT_STATE.yaml` — usually at the manuscript's root.

Read it. document_phase.current gates invocation; meta.guidance_level controls framing; meta.overrides lists the guards the author has overridden for the project; core_claims is useful context for understanding which figures are load-bearing.

Figure locations are usually NOT declared in `MANUSCRIPT_STATE.yaml`. The schema does not require a figure index. Discover figures from the manuscript text itself: caption blocks (commonly introduced by **Figure N.**, Figure N:, Fig. N., or a Quarto #| fig-cap:), and body-text references (Figure N, Fig. N, Fig N, figs. N–M). If the manuscript declares figures more structurally (e.g. Quarto #| label: fig-* with cross-references), prefer that. If neither captions nor references can be located, stop and tell the author the skill found no figures to align.

If MANUSCRIPT_STATE.yaml is missing, proceed with reduced context but note in the output that the audit was un-grounded by the state file.

Conversational style

Read meta.guidance_level from MANUSCRIPT_STATE.yaml (default standard if absent). Adapt framing — not the structured output — per [[guidance-level]]:

• terse — open with a one-line "running figure-text alignment

(text-only)"; emit the markdown report; no closing summary.

• standard — open with a sentence naming the manuscript and the

number of figures discovered; close with a one-line summary of the findings.

• full — open with what this skill produces (per-figure

caption-vs-body-text alignment classification + pattern-level flags) and what it explicitly does not do (read figure images); close with which findings to act on first and which are informational. If running for the first time in this session, also offer /scriptorium:explain figure-text-alignment so the author can learn the skill's design before reading its output.

Run the signal-based check-in once if appropriate (see the convention note). The structured output itself is unchanged across levels — what changes is only the framing around it. Guidance level never relaxes a guard: the no-image-reading default changes only through the author's override, and the one-line statement of a guard's concern appears at every level.

Operational protocol

Work in this order. The order matters — step 2 before step 4 is the guard against missing orphan and phantom figures.

1. Read `MANUSCRIPT_STATE.yaml`. Extract:

• document_phase.current — if outline, apply the

outline-phase guard (see Guards and author overrides).

• meta.guidance_level — framing only; see above.
• meta.overrides — guards the author has overridden for this

project; apply them in the precedence above.

• core_claims — useful context for which figures are

load-bearing, even if the field is not load-bearing itself.

1. Discover figures from the manuscript text. For each figure,

record:

• The figure ID (Figure 1, Figure 2A, Fig. 3, etc.).
• The caption text, verbatim.
• The panel labels declared in the caption (A, B, C, …), if any.
• Every body-text sentence (or clause) that references this

figure, with its location (section, paragraph or line).

1. Build the cross-reference inventory. Two sets:

• Figures that have a caption.
• Figure IDs referenced in the body text.

The set difference is where orphan and phantom flags come from.

1. **For each figure with both a caption and at least one body-text

reference**, walk through these steps (mirroring the citation-audit four-step pattern):

1. Extract the caption's claim: what does the caption say

the figure shows? Note panel structure if any.

1. Extract each body-text reference's claim: what does the

sentence assert the figure shows? Note panel reference if any.

1. Compare the two on three axes:

• Subject — same variable / dataset / comparison?
• Direction / pattern — does the body text describe an

increase / decrease / no-difference that the caption also names (or contradicts)?

• Panel and axis specifics — does the body text point at a

panel that the caption defines? Are units / log-vs-linear / raw-vs-normalised consistent?

1. Classify the alignment as one of:

• Aligned — caption claim and body-text claim describe

the same content / pattern / direction.

• Partially aligned — overlapping but with a meaningful

divergence (different panel referenced, different axis named, different direction implied for a sub-claim).

• Misaligned — caption and body text disagree about what

the figure shows.

• Cannot determine — caption is too sparse to compare,

or the body-text reference is too vague (e.g. a bare "see Figure 3" with no claim).

1. Scan for pattern-level flags independent of the per-pair

alignment classification:

• Orphan figure — figure exists (has a caption) but is

never referenced in body text.

• Phantom figure — body text references "Figure N" but no

caption for Figure N exists.

• Panel mismatch — caption describes panels A/B/C; body

text references a panel letter the caption does not define (e.g. body text says "Figure 2D" but the Figure 2 caption defines only A/B/C).

• Axis / units divergence — caption names units / scaling

("log₁₀ counts", "fold-change") that the body text discusses in incompatible terms ("raw counts", "absolute difference").

• Direction divergence — caption says one direction

("decrease", "downregulation"); body text discussion of that figure asserts the opposite ("increase", "upregulation").

1. Emit the report. Use the section headings below verbatim so

downstream skills and future orchestrators can consume the output by structure.

Output format

Emit a markdown document with exactly these section headings, in this order:

# Figure-text alignment (text-only)
 
Overrides applied: <guard ids>   (only when the run proceeded past a scope guard)
 
## Summary
 
-Figures discovered (caption present): N
-Figure IDs referenced in body text: M
-Per-pair alignment:
  -Aligned: A | Partially aligned: B | Misaligned: C | Cannot determine: D
-Pattern flags:
  -Orphan figures: E
  -Phantom figure references: F
  -Panel mismatches: G
  -Axis / units divergences: H
  -Direction divergences: I
 
## Per-figure assessment
 
| Figure | Caption excerpt | Body-text reference excerpt | Alignment | Notes |
|---|---|---|---|---|
 
(One row per (figure, body-text reference) pair. A figure referenced
in three different paragraphs produces three rows. Excerpts are
short — 10-20 words. "Notes" is one sentence: what the assessment
hinges on. Figures that have a caption but no body-text reference
appear under Pattern flags → Orphan figures, not here. Under the
`figure-image-reading` override, "Notes" starts with
`unvalidated vision` for any image-derived finding. Under the
`critique-rewrite-suggestions` override, a flagged row or pattern
flag ends with one `Suggestion:` line.)
 
## Pattern flags
 
(One subsection per pattern type that turned up. Omit empty
subsections.)
 
### Orphan figures
-Figure N — caption present (section / location), but no body-text
  reference found.
 
### Phantom figure references
-"Figure N" referenced at (section / location), but no caption for
  Figure N was discovered.
 
(Under the `generation-from-blankness` override, a `Draft:` line
follows an Orphan or Phantom flag: a body-text reference for an
orphan, a caption for a phantom, with placeholders bracketed.)
 
### Panel mismatches
-Figure N caption defines panels {A, B, C}; body text at (location)
  references "Figure NX" where X ∉ {A, B, C}.
 
### Axis / units divergences
-Figure N caption uses "<unit-or-scaling>"; body text at (location)
  discusses the same figure in terms of "<incompatible unit / scaling>".
 
### Direction divergences
-Figure N caption asserts "<direction>"; body text at (location)
  asserts "<opposite-direction>" of the same comparison.
 
## What this skill did NOT check
 
(Honest list. Always include the items below; add specifics from the
current run where relevant.)
 
-**Did not read figure images.** This is the text-only subset of
  figure-text alignment. The multimodal counterpart (sub-skill B) is
  the right skill for that and is deferred until LLM-vision
  reliability for scientific figures is validated against a
  known-mismatch test set (see roadmap v0.3 deferred section, issue
  #14). Until sub-skill B ships, no claim in this report rests on
  what the figure actually displays — only on what its caption text
  says it displays. (Under the `figure-image-reading` override,
  replace this item with: "Read figure images under an author
  override; every image-derived finding is labelled `unvalidated
  vision`." The image-dependent items below (axis labels, rendered
  annotations, pattern-claim verification) then state what was
  actually read, labelled `unvalidated vision`, and the title drops
  "(text-only)".)
-Whether the figure's actual axis labels match what the caption
  claims they are. Requires reading the image.
-Whether statistical annotations rendered on the figure (asterisks,
  p-value text, error-bar style) are consistent with statistics
  reported in the text. Requires reading the image and is also
  partly the territory of a future `statistics-consistency` skill.
-Image integrity (duplication, manipulation, splicing). This is
  emphatically out of scope and belongs to Proofig / ImageTwin /
  human inspection — see [[forensic-methodology]] for the boundary.
-Whether the figure is the *right* figure to support the body-text
  claim. The skill audits alignment, not editorial judgement about
  figure choice.
-Whether the caption itself is well-written (style, length,
  completeness against journal guidelines). Caption-quality is
  separate from caption-vs-text alignment.
-Pattern-claim verification — "is the trend the author describes
  actually visible in the figure?" Requires reading the figure and
  often the underlying data.
-Sample-size consistency across the manuscript (Methods N vs.
  Results N vs. caption n=). That is internal-consistency work
  closer to a future `statistics-consistency` skill; this skill
  flags axis / unit divergences but not numeric-N drift across
  sections.

What "good output" looks like

• Specific, location-anchored. Never "some figures appear

misaligned." Always "Figure 2 caption says panel A shows downregulation; body text at Discussion ¶3 says Figure 2A shows upregulation."

• Conservative under uncertainty. When the caption is too sparse

or the body-text reference too vague, mark cannot determine and explain why. Do not guess.

• Quantitative summary at the top. The Summary section is what a

busy author scans first; pattern-flag counts let them triage.

• Patterns over enumeration. If a single phantom figure is

referenced in twelve places, it is one phantom-figure pattern row with twelve locations, not twelve separate rows.

• Honest scope statement. Every report includes the *did NOT

check* list, naming the multimodal deferral by name. Authors must not mistake the text-only audit for a full figure-text-alignment pass.

What you must not do

• Read or interpret figure images, screenshots, or rendered plots,

unless the author overrides figure-image-reading.

• Invent figure content. If the caption does not state what the

figure shows, do not speculate. (no-invented-manuscript-facts; fixed.)

• Modify the manuscript, the captions, or the figure files.
• Suggest specific rewrites of captions or body-text references,

unless the author overrides critique-rewrite-suggestions. By default, flag the misalignment; the author decides what to change.

• Score the manuscript on a quality scale. Audit is descriptive,

not evaluative.

• Conflate this skill with image-forensics work. Image integrity

(Bik-style duplication / manipulation detection) is a different problem with different methodology and lives outside scriptorium — see [[forensic-methodology]]. (no-image-integrity-claims; fixed.)

• Operate on outline-phase manuscripts, unless the author overrides

outline-phase. By default, state the concern in one line, decline, and tell the author how to override.

Grounding

This skill is grounded in scriptorium's knowledge layer:

• [[visualization-figures]] — primary grounding. Names figure-text

alignment as a documented manuscript failure mode (wrong panel referenced, axis-unit mismatch, "trends" not visible in the cited figure, figure-counter drift). Anchors the text-only / multimodal split this skill embodies: cross-reference, panel-letter, and counter-drift checks are text-tractable; axis-label match and plot-type match are not, and are deferred to sub-skill B until LLM-vision reliability on scientific figures is validated. The 2025–2026 benchmarks put numbers on that deferral: on PRISMM-Bench, 21 large multimodal models scored 27.8–53.9% on real reviewer-flagged inconsistencies, and on MuSciClaims (a preprint) the best figure-claim verification result was 0.72 F1, with models biased toward judging claims as supported — the failure that would let an image-reading sub-skill miss real misalignments while appearing to pass them.

• [[internal-consistency]] — frames figure-text alignment as a

class of internal-consistency failure (alongside terminology drift, numerical-claim consistency, methods–results–discussion alignment). Provides the structured-output discipline that each flagged discrepancy emits enough location information for the author to navigate to both passages.

• [[forensic-methodology]] — used here for the boundary

statement. Bik-style image forensics is a different problem (figure integrity, not figure-text alignment) with different methodology (image processing, corpus comparison) and is emphatically out of scope for scriptorium. Naming the boundary in What this skill did NOT check prevents authors from mistaking a text-only alignment audit for an integrity audit. The 2025–2026 evidence strengthens the exclusion: PhD-level reviewers spotted AI-forged western blots and tumour images at chance (50.5%), and the best commercial detector reached AUC 0.79. Even sub-skill B, when it ships, should not take on image-integrity claims. When pointing authors to ImageTwin, note that in ASM's 2023–24 pilot it flagged problems only in halftone images (blots, micrographs, photographs), not graphs or line art. No source we found states Proofig's image-type coverage, so do not extend that limit to Proofig.

• [[guidance-level]] — scriptorium-wide convention controlling

how much framing the skill adds around its structured output.

• [[declared-work-scope]] — scriptorium-wide convention. This

skill operates on declared work: figure captions the author has written and body-text figure references the author has placed. By default it does not generate captions, does not invent body-text references, and does not propose figures the manuscript should add (the generation-from-blankness and outline-phase scope guards).

• [[author-authority]] — scriptorium-wide convention for guards.

The scope guards above are defaults the author can override, per run or under meta.overrides; the accuracy guards are fixed. It also defines the "Overrides applied" line.

• [[copyediting-vs-developmental]] — source of the

critique-rewrite-suggestions default: a critique skill queries problems for the author to fix rather than rewriting the text.

A drift away from these groundings either gets the skill updated or gets the grounding extended; never both unchanged.