Research Radar — Methodology

01. Confidence Shifts Within a Topic

Groups published briefs by topic tag, sorts each group chronologically, and flags a pair when a later brief sits higher on the evidence ladder (speculative → contested → emerging → established) than an earlier one, at least 7 days apart — a same-day gap between two takes on one story is treated as noise, not a shift.

What this is not: a topic tag is not a mechanism or a single study. Two briefs in a pair are frequently different findings that happen to share a topic, not the same claim being re-confirmed over time. We tested a stricter version of this mechanism — grouping by shared historical antecedent instead of topic, so a pair only counts if it's plausibly the same underlying mechanism maturing — and it currently returns zero hits against the live archive. That version ships once the archive is deep enough to support it; until then, this section is reported as topic-level confidence movement, not mechanism maturation.

evidence_position is classified per brief against the rubric on our editorial policy page, not left at a schema default.

02. Cross-Framework Convergence

Requires two or more distinct categories, computed from autoresearch_antecedents — historical scientists whose work a current finding extends or validates, identified per brief. An earlier version of the prompt suggested a handful of named scientists as examples; it visibly biased the output toward them and was removed 2026-08-09.

When two scientists' convergence lists overlap heavily (e.g. co-cited on most of the same briefs), that's noted inline rather than presented as two independent signals.

03. Trial Registry Activity (prototype)

Institutional and clinical buyers diligence things a citation graph doesn't show — whether a finding already has trial-registry or IP activity behind it. For each brief, an LLM pass names the single most specific compound, drug, or intervention class the finding is actually about (or returns nothing, if it's a bench-level mechanism, gene, or pathway with no named intervention a trial would list), then checks that term against ClinicalTrials.gov's public API.

Specific vs. generic: a term that returns more than 150 total trials is excluded from the signal list — at that volume, `query.intr` is matching so loosely (e.g. "exercise," "statin," "omega-3 fatty acids") that it isn't telling you anything about this specific finding. The 150 cutoff was picked empirically: named compounds we know are specific (senolytic, urolithin A, a named vaccine adjuvant) land in the tens; generic terms land in the hundreds to tens of thousands.

Coverage, stated plainly: roughly 15-20% of the archive currently names something specific enough to check. Most briefs describe a mechanism, gene, or pathway with no compound named yet — that's reported as zero matches, not papered over with a weak one. This is U.S.-only (ClinicalTrials.gov); European (EU CTIS) and Asian (WHO ICTRP, ChiCTR, jRCT) registries were evaluated and are real but meaningfully harder to reach live — WHO ICTRP is a weekly bulk export, not a queryable API, and EU CTIS requires registered API access. Those are a future layer, not this one.

Scope

All mechanisms run against the full archive, published briefs only — draft content is excluded.

Changelog

• 2026-08-10 — Added mechanism 03 (trial registry activity, prototype). Fixed a date-arithmetic bug where full-timestamp subtraction (rather than calendar-date subtraction) floor-divided some gaps short by a day. Reframed Confidence Shifts from "evidence maturation" to make the topic-vs-mechanism distinction explicit.
• 2026-08-09 — Removed named example scientists from the antecedent-finding prompt after A/B testing showed it biased output toward those examples.