Docs
This is the reference for the two command-line tools that produce this corpus. It documents what you type. What the resulting graph then knows — what a layer is, why staleness is computed rather than remembered, what an approval attaches to — is on the pipeline pages, where every figure is generated from the corpus rather than written down.
The division is not editorial. These pages describe commands, which change when the code changes; those pages describe a model, which changes when a decision changes. Keeping them apart is what stopped this section describing a system that no longer existed.
The two tools
scripts/pipeline.py is the pipeline. It reads pipeline/layers.yaml, works out what the
layer you asked for still needs, runs each one, and appends a record to
runs/<paper>/ledger.jsonl naming every input by path and content hash. It is the only thing
that writes a ledger, which is what makes "is this result still current" a question about
files rather than about anyone's memory.
elife-extract is a set of layer runners — one subcommand per node in the graph that a
model or a script can execute. Each reads the paths its layer declares and writes the path its
layer declares. You can run them directly; nothing then records that you did.
That asymmetry is the whole reason to prefer the first. There is no second path that produces the corpus.
python3 scripts/pipeline.py run gadeke-2026-guilt-insula claim-tree --dry-runprepare cd extract && python3 -m elife_extract.cli prepare --paper gadeke-2026-guilt-insula …
results-reader cd extract && python3 -m elife_extract.cli results-reader --paper gadeke-2026-guilt-insula
caption-reader …
structure-reader …
reconcile …
external-review …
edge-inference …
claim-tree …
Eight commands, in dependency order, none of them run. Every one of them is a layer with a page of its own.
Where to start
- Install and configure — Python, credentials, the environment-variable contract.
- One paper end to end — the walkthrough, from a DOI to a claim tree with a coverage report.
- The runner —
pipeline.py:graph,backfill,state,run,approve. - The layer runners —
elife-extract, one subcommand per layer. - Batch operation — running the corpus, and what it costs.
- Measuring a prompt change —
evaluate, the one subcommand that is not a layer. - Configuration and the prompts — the reference tables.
- For contributors — code structure, adding a prompt variant, what the tests pin.
Coming from the old CLI
The command surface changed. If you have older notes or scripts:
| If you used | Use now |
|---|---|
elife-extract extract --doi <doi> |
pipeline.py run <paper> claim-tree, or prepare then each reader |
elife-extract run |
pipeline.py run <paper> claim-tree |
write --draft out/draft-<slug>.json |
write --paper <slug> |
write --review-mode external |
the external-review layer, which runs in the chain |
write --review-mode interactive |
pipeline.py approve <paper> <layer> --by NAME, after the write |
write --review-mode auto-approve |
nothing — there is no gate to bypass |
coverage --doi <doi> --claims-dir <dir> |
coverage --paper <slug> |
mark --doi <doi> … |
mark --paper <slug> |
scripts/verify-references.py |
pipeline.py run <paper> reference-check |
Elsewhere on this site
| Papers | the claim graphs themselves |
| Pipeline | the layer graph, and which papers have been through each layer |
| Roles and relations | what a claim can be, and what can hold between two |
| Methodology | the procedure the corpus follows |
| Design note | why the system is shaped this way |