The intake pipeline

How a paper travels from a source feed to a graded record, which model does which step, and what gets excluded.

Every paper on this site arrives the same way. Nothing is added by hand. This page describes the steps, which model does which one, and what gets thrown away.

The steps

1. Fetch. Source feeds are searched daily for new work matching the tracked topics. See sources for which feeds and which topics.
2. Get the full text. Where an open-access PDF exists it is downloaded and parsed. Where it does not, the abstract is all the later steps have to work with, and that limits how much can be said about a paper.
3. Summarise. A model writes the one-sentence summary and the headline used in listings.
4. Assess and extract claims. A model assigns study design, evidence strength, methods rigour and a relevance score, then pulls out individual findings as claims with their effect sizes and sample sizes.
5. Apply the relevance gate. Papers scoring below 7 are removed from the corpus; a record of each exclusion, with its abstract, assessment and extracted claims, is kept so the decision stays auditable.
6. Categorise claims. Each claim is mapped to a canonical outcome category so claims about the same thing can be compared across papers.
7. Detect tensions. A SQL rule, not a model, flags pairs of claims in the same outcome category that point in opposite directions. See the evidence layer for what that does and does not establish.
8. Synthesise. Where a category has accumulated enough evidence, a stronger model writes a state-of-the-evidence page.

Why the relevance gate is strict

The gate is the only thing standing between a topical corpus and a general paper dump. Search feeds return a great deal that is adjacent but not relevant, and a corpus that accepts everything stops being useful for the question it exists to answer.

Exclusion is not deletion of the reasoning. The exclusion record keeps what the model saw and what it concluded, so a decision can be revisited rather than merely trusted.

Which model does what, and why

Model choice here is mostly an economics problem. The pipeline runs every day over a large volume of papers, so per-paper work has to be cheap, while the handful of steps that produce something a person reads end to end can afford a stronger model.

High-volume work

Summaries, assessments and claim extraction run on a small fast model. This is the bulk of the spend and the bulk of the output.

Synthesis

A stronger model, used only for the written state-of-the-evidence pages, where holding many claims in view at once actually matters.

Tensions and links

Plain SQL. A rule that compares directions within a category is auditable and reproducible in a way a model's judgement is not, and this is a place where being predictable beats being clever.

Live counts for models, costs and throughput are on the methods page rather than here, so this page does not go stale between runs.

Where this pipeline fails

• A paper with no accessible full text is assessed from its abstract, which systematically understates what a careful study did.
• Relevance scoring is a model judgement. A borderline paper can be excluded for reasons a human would disagree with.
• Claim extraction reports what a paper says, not whether it is right. A confidently stated result from a weak study still becomes a claim, which is why the evidence grade attached to it matters.

The full list is on known limitations.