The evidence layer
How claims are graded, what each label means, and how tensions between claims are detected.
A claim is a single finding pulled out of a paper, with its direction, its effect size where one is reported, and a grade for how much weight it can carry. Claims are what makes findings comparable across papers, and the grading is what stops that comparison from being misleading.
Evidence grades
Each claim carries a grade describing the strength of the support behind it, not whether it is true:
high- Clean identification and a design that can support a causal reading.
medium- Credible, with real caveats about design, sample or generalisability.
low- Correlational or descriptive; consistent with the claim but far from establishing it.
speculative- Argued rather than measured, or measured in a way that cannot bear weight.
Grades are assigned by a language model from the paper's own description of its design. A paper that oversells its identification strategy can be graded too generously.
Direction
Direction records which way the effect ran: positive,
negative, mixed, or null_result.
null_result is a real finding, not a missing one, and it is worth
filtering for on its own: an absence of effect is frequently the most useful
thing in a literature dominated by positive results.
Outcome categories
Each claim is filed under a canonical outcome category so that claims about the same underlying thing can be compared even when the papers use different vocabulary. This is what makes Explore possible, and it is also the main place where a mis-categorisation quietly distorts a comparison.
Tensions
Where two claims in the same outcome category point in opposite directions, they are flagged as being in tension. This is a SQL rule, not a model judgement, which makes it predictable and reproducible.
A tension is not a confirmed contradiction. The rule compares direction within a category. It cannot tell whether the two studies share a population, a time period, a treatment or an outcome definition. Two studies can honestly disagree in sign and both be right about different things. Read a tension as an invitation to open both papers.
Reading the labels in the digest
Entries in the weekly digest carry two independent tags and, in parentheses, the study design. One tag says how the paper moves the existing picture; the other says how strong the support is. They are deliberately kept apart, because a strong study can be unsurprising and a weak one can be novel.
Relation — how it moves the picture
- New
- A finding on a question the standing evidence had not settled. Also used for capability benchmarks and theory, which inform an economic-outcome view but cannot confirm one.
- Confirms
- Same outcome, same direction, comparable population as the standing view. An independent study that raises confidence.
- Extends
- Same direction, but a new setting, population or mechanism.
- Tension
- A direction or magnitude mismatch, not yet strong enough to overturn anything.
- Challenges
- Reserved for a same-estimand finding whose identification is at least as strong as the standing base. A single study does not flip a verdict.
Evidence status — how strong the support is
- Established
- Multiple well-identified studies point the same way.
- Suggestive
- Real but limited: one study, or several with caveats.
- Framework
- A theoretical model or argument, not an estimate.
- Descriptive
- Measures or documents something without identifying a causal effect.
The study design in parentheses (RCT, quasi-experiment, panel, descriptive) is a separate fact about how the study was run, and is kept apart from both tags on purpose.
Querying the evidence layer
Everything above is exposed over the API. search_claims filters by
direction, outcome category, evidence grade, study design and effect type. See
the tool reference.