Skip to content

Glossary

This glossary defines terms that appear across the documentation. Where a term has a dedicated page, the glossary entry links to it. Where the term is defined in the methodology document, the glossary cites § X of docs/method.md in the corpus repo.

Terms

Section titled “Terms”

Claim. One declarative proposition about the paper’s content, anchored to a specific panel (when panel-grounded) or to the paper’s argument structure (when hypothesis-, prediction-, or synthesis-level). The atomic unit of the elife-claim-trees schema. See the claim schema and docs/method.md § 4.

Claim-type. The epistemic character of the proposition itself — what kind of knowledge claim it is. One of empirical, interpretive, existence, synthesis, assessment (the methodology specification) or extended to include hypothesis, prediction, scope, methodological per actual corpus usage. Distinct from role.

Curated reference / curated corpus. The 12-paper claim graph at ~/Projects/mainenlab/elife-claim-trees/claims/ extracted by hand by the curator. The gold standard for round-trip validation.

Deductive layer. The implicit hypothesis-prediction-test scaffolding that the curator infers from a paper’s modelling decisions but which often isn’t explicitly stated in the prose. The structural-inference work that the external reviewer (Step 4.5) tries to recover.

Draft / draft claim table. The output of Step 4 (reconciliation). A DraftClaimTable JSON with the reconciled list of claims, confidence-tagged, with per-agent evidence quotes preserved. The input to Step 5 (review).

Edge. A typed proposition about logical structure between two claims. 14 edge types defined; see the 14 edge types for the full inventory.

External reviewer. Step 4.5 of the system’s pipeline (an extension to the methodology). An Opus call that reads the paper plus the reconciled draft and produces a revised draft addressing seven systematic biases. Substitutes for the analyst at the review gate when the CLI runs without a curator. See the external reviewer.

HAAK. The research platform this system is built within. Provides the model-router infrastructure, the Vertex AI configuration, and the conventions the CLI follows. Not user-facing for eLife collaborators.

Paper-slug. The directory name and identity key for a paper in the corpus. Lowercase, hyphenated, derived from <author>-<year>-<title-keywords> (e.g., headley-2026-inhibitory-rhythms). Override via --paper-slug.

Panel. A figure panel ID anchoring a claim to where its data lives (e.g., fig4a, fig5d-h, figS2c). Required for empirical claims; null for hypothesis / synthesis / scope-level claims.

PreparedPaper. The output of Step 1 (prepare.py). A dataclass with the paper’s slices (abstract, results, captions, methods) ready for the three extraction agents.

Reconciliation. Step 4 of the methodology. Folding three independent agent extractions into one confidence-tagged draft claim table via an Opus call. See reconciliation.

Recovery. A round-trip metric. (# reference claims with a CLI match) / (# reference claims). The load-bearing acceptance metric — measures whether the CLI is finding what the curator found.

Reference / reference corpus. Same as curated reference. The gold standard claim files used as the comparison baseline in evaluate.

Review gate. Step 5 of the methodology. The hard rule that no claim files are written until the draft is approved. Implemented by the CLI’s four --review-mode options. See the review gate.

Review modes. Four options for the write subcommand’s --review-mode flag: interactive (analyst edits in $EDITOR), external (Opus reviewer pass substitutes for analyst), auto-approve (no review), dry-run (print only). See review modes — when to use each.

Role. The rhetorical function a claim serves in the paper’s argument. One of 9 values: hypothesis, prediction, empirical, control, scope, methodological, synthesis, interpretation, literature-context. Distinct from claim-type. See the 9 roles.

Round-trip. Validation methodology. Run the pipeline on a paper that has a curated reference; score the CLI’s output against the reference. Implemented by the evaluate subcommand. See how we measure quality.

Slug. A filename-safe identifier — lowercase, hyphenated, no special characters. Used for paper-slugs (directory names) and claim-slugs (file names within a paper directory).

Structural inference. The reasoning step that recognizes a paper’s hypothesis-prediction-test structure when the prose flattens it into a sequence of empirical findings. The work the external reviewer tries to do that prose-level extraction systematically misses.

Three-agent partition. Steps 2-3 of the methodology. Three independent extraction agents (Results-reader, Caption-reader, Structure-reader) each read a partitioned slice of the paper. See the three-agent partition.

Verify-refs. A CLI subcommand that resolves DOIs for role: literature-context claims via CrossRef. Anti-hallucination check on existing DOIs; resolution from Author (Year) hints when DOIs are missing. See the five subcommands.

Acronyms and shorthand

Section titled “Acronyms and shorthand”
  • DOI — Digital Object Identifier, the canonical paper identifier (e.g., 10.7554/eLife.95562)
  • MDX — Markdown with JSX components, the format Astro uses for documentation
  • PDF — Portable Document Format, the input the system fetches from the eLife CDN
  • JSON — JavaScript Object Notation, the wire format between pipeline steps
  • YAML — the format used for claim file frontmatter
  • CrossRef — the citation registry queried by verify-refs to confirm cited DOIs
  • Vertex AI — Google Cloud’s managed-LLM service, the canonical Anthropic backend in HAAK