Overview
Five-minute orientation. What the system does, how it relates to the eLife claim-trees methodology, when to use it.
eLife Claim Trees — Documentation
An eight-step claim induction pipeline. Three Sonnet readers, an Opus reconciliation, and an Opus reviewer that recovers the deductive structure prose extraction misses. Built to be operated by your team, not just demonstrated.
Where to start
Section titled “Where to start”Architecture
Pipeline diagram, the three-agent partition, reconciliation, the external reviewer pass, and the review gate.
Using the CLI
Install, configure, run on your first paper, scale to your corpus. Cost and performance characteristics.
Validation
Round-trip methodology. Per-paper and aggregate scorecards from the 10-paper public eLife corpus.
Reference
Configuration reference, the 9 roles, the 14 edge types, the prompts, the API.
For contributors
Code structure, prompt-variant authoring, validation discipline, design decisions.
What this is
Section titled “What this is”A working implementation of the eight-step claim-induction methodology described in docs/method.md. Given an eLife paper DOI, the system fetches the paper, runs three independent extraction agents on partitioned slices of the prose, reconciles the agents into a draft claim table, optionally runs an Opus structural-inference pass to recover the deductive layer, and writes claim files in the corpus schema — with CrossRef DOI verification for cited references.
It is built as a Python CLI (elife-extract) with a stable command surface, a Pydantic-validated wire format, and a round-trip evaluation harness for measuring the impact of any change against the curated reference corpus. Every architectural decision is documented; every prompt is in version control; every empirical claim about performance is backed by a per-paper scorecard you can re-run.
Audience
Section titled “Audience”This documentation is for the people who will operate the system, integrate it with eLife’s editorial infrastructure, and extend it as the methodology evolves. It assumes Python familiarity, a working knowledge of the LLM tooling stack (Anthropic via Vertex AI), and access to the elife-claim-trees corpus or your own corpus following the same schema.
If you are looking for the methodology rather than the implementation, see docs/method.md in the corpus repo. If you are looking for the per-paper claim graphs, see /papers/.