The claim schema

The claim is the system’s atomic unit. Every step of the pipeline either produces or consumes claims. This page describes the schema — what fields a claim carries, what each field means, and how a claim file looks on disk.

The canonical schema specification is docs/method.md § 4 in the corpus repo. This page summarizes; for edge cases consult the methodology directly.

The claim, conceptually

A claim is one declarative proposition about the paper’s content, anchored to a specific panel (when the proposition is panel-grounded) or to the paper’s argument structure (when the proposition is hypothesis-, prediction-, or synthesis-level). Each claim is one file; the file’s frontmatter encodes the structured fields, the body holds prose elaboration.

A claim is not a sentence quoted from the paper. It is a proposition extracted from the paper, phrased as a declarative sentence, with provenance back to the source text via verbatim quotes in the body or via the assertions block.

Required fields

Every claim has these fields in its YAML frontmatter:

Field	Type	Meaning
`uuid`	UUID4	Generated once at creation, immutable. The claim’s stable identity across renames, splits, merges.
`slug`	string	Filename slug. Lowercase, hyphenated, 3-6 words, verb-phrase form (e.g., `distal-inhib-drops-firing-02hz`).
`doi`	string or `~`	For literature-context claims, the cited paper’s DOI. For other roles, `~` (claims aren’t yet citable units).
`claim`	string	The declarative sentence stating the proposition. Active voice. Quantitative where the result is quantitative.
`claim-type`	enum	The epistemic character of the proposition. One of `empirical`, `interpretive`, `existence`, `synthesis`, `assessment`.
`role`	enum	The rhetorical function the claim serves in the paper’s argument. One of 9 values (see the 9 roles).
`concepts`	list	Controlled domain terms (e.g., `[dopamine-transporter, clearance, dorsal-striatum]`). Filled at review.
`priority`	date	When the claim was first registered. Set to today’s date at write time.
`epistemic`	enum	`strong` / `moderate` / `weak` / `contested` — the analyst’s overall assessment of evidentiary support.

Optional but conventional fields

Field	Type	Meaning
`displayClaim`	string	A 1-2 sentence form for body-text rendering. Softer than the formal `claim` field.
`shortClaim`	string	A short clause (≤90 chars) for graph nodes and tooltips.
`belongings`	list	General typed edges to other claims (`{relation, target}` pairs). Distinct from the named edge keys below.

Edge fields (typed relationships between claims)

Edges are top-level keys whose value is a list of target claim slugs. The methodology defines 14 edge types — see the 14 edge types for the full list. The most common:

Edge	Meaning
`entails`	Hypothesis-to-prediction deduction. A hypothesis carries `entails:` to its predictions.
`derived-from`	Reciprocal of `entails`. A prediction carries `derived-from:` back to the hypothesis.
`tests`	An empirical claim that tests a prediction. Closes the hypothesis-prediction-test loop.
`requires`	A’s validity depends on B’s. A mechanistic or hierarchical dependency.
`supports`	A provides evidence for B. The standard inductive-support edge.
`dissociates-with`	Symmetric. Two empirical claims that jointly establish a contrast.
`rules-out`	A’s evidence eliminates an alternative explanation B.
`interprets`	A reframes empirical B through a theoretical lens.
`scopes`	A bounds the interpretation of B (or `[*]` for global scope).

Provenance fields

Field	Type	Meaning
`assertions`	list of blocks	Each block links the claim to a specific paper-panel-analysis tuple. Required field — every claim has at least one assertion (the paper that asserts it).
`reproductions`	list of blocks	Each block records a verification attempt: which agent ran it, when, what status (`verified`, `unverified`, `failed:mismatch`, etc.), pointers to the script and reproduced figure. Empty by default; populated by per-paper `verify.py` scripts.

A real claim file

Here is distal-inhib-drops-firing-02hz.md from the curated Headley reference, abridged:

---
uuid: 26819b09-5b20-4c90-b9f3-8bdd29a2a57c
slug: distal-inhib-drops-firing-02hz
doi: ~
claim: >
  Doubling the strength of distal dendritic inhibition reduces somatic firing rate
  from a baseline of approximately 5.5 Hz to approximately 0.2 Hz, primarily by
  suppressing the occurrence of dendritic Ca²⁺ and NMDA spikes rather than by
  directly raising AP threshold.
displayClaim: >
  Doubling distal dendritic inhibition collapses somatic firing from ~5.5 Hz to
  ~0.2 Hz, almost entirely by suppressing dendritic Ca²⁺ and NMDA spikes.
claim-type: empirical
role: empirical
concepts:
  - distal dendritic inhibition
  - somatic firing rate
  - dendritic spike suppression
  - Ca2+ spikes
  - NMDA spikes
priority: 2026-03-30
epistemic: strong

tests:
  - prediction-distal-dendritic-spike-mechanism

dissociates-with:
  - perisomatic-inhib-drops-firing-07hz

belongings:
  - relation: requires
    target: l5-model-single-cell-scope
  - relation: supports
    target: hypothesis-distinct-compartmental-roles

assertions:
  - paper-slug: headley-2026-inhibitory-rhythms
    doi: 10.7554/eLife.95562
    panel: fig4, fig5
    figureUri: https://iiif.elifesciences.org/lax/95562%2Felife-95562-fig4-v1.tif/...
    analysis: scripts/Fig4.ipynb, scripts/Fig5.ipynb
    dataset: https://datadryad.org/dataset/doi:10.5061/dryad.v6wwpzhb8
    method: compartmental modelling — inhibition magnitude sweep
    confidence: strong

reproductions:
  - agent: mainen-z
    date: 2026-03-30
    status: verified
    script: verification/headley-2026-inhibitory-rhythms/verify.py
    original_figure: verification/originals/.../fig4.jpg
    figure: verification/headley-2026-inhibitory-rhythms/fig4a-firing-rates.png
    notes: >
      Verified directly from Figure4a.csv in repo data/. Pre-computed firing rates
      (30 trials each condition): control=5.5±0.86 Hz, dendritic=0.2±0.15 Hz...
---

The near-complete suppression of firing (5.5 → 0.2 Hz) under doubled distal
inhibition demonstrates that dendritic spikes are not merely modulatory — they
are required for most somatic APs under these conditions. The mechanism is
dendritic spike suppression: distal inhibitory conductance prevents membrane
voltage from reaching spike threshold in the apical tuft...

What the system produces vs the curated reference

The CLI emits claims that conform to this schema, but with a few characteristic differences from the curated reference:

Empty edge sections. Step 6 (dependency mapping) is scaffolded; claim files emit with belongings: [] and no top-level edge keys. The analyst populates these at review.
Empty reproductions. Step 8 verification is per-paper Python; the CLI doesn’t run it.
Auxiliary fields (displayClaim, shortClaim) absent. The agents don’t author these; the analyst adds them at review.
epistemic: tentative as default. The CLI’s agents emit confidence: tentative for their own assessment; the curator’s epistemic: strong/moderate/weak/contested is an analyst judgment that the CLI doesn’t attempt.

These are the places where the analyst’s review value is highest. The CLI does the mechanical work (extraction, role classification, panel assignment, evidence quoting); the analyst does the semantic work (edge mapping, epistemic assessment, auxiliary phrasings).

Where to read more

docs/method.md § 4 in the corpus repo — the canonical schema specification
The 9 roles — each role’s definition, signal phrases, and edge patterns
The 14 edge types — the full edge inventory with reasoning forms
What the curator does, what the system does — the human/system division of labor