Two surfaces accompany the Frame Check product: the library of named frame patterns, and the methodology that documents how detection and verification work.
16 named frame patterns with identification cues, generation affordances, and worked examples. Curated and versioned. Licensed CC-BY-4.0.
How Frame Check measures structural framing, which sources are consulted for verification, calibration evidence, and honest limitations.
Methodology for the five-dimension decision-support profile. Strategic flag for Track C; profile pending validation. Open for pre-implementation review.
A snapshot of how frontier AI models framed a fixed set of factual questions in the observed window, with per-topic voice and verdict breakdown. Slice, not live dashboard.
7 calibration runs measuring per-provider precision, recall, and F1 against a seed corpus of claims with independently established ground truth.
Expert raters needed to validate the decision-readiness profile against rated documents. 30-60 min per document, 5-10 doc batches, attribution + compensation options. Without raters the profile stays experimental indefinitely.
10 corpus entries the Phase 2 validation harness consumes: per-entry document, computed Frame Check profile, diff/peer artifacts. Aggregate findings cite these documents; surfacing the corpus lets readers verify findings against the actual text.
Pedagogical readings of corpus diff and peer artifacts, plus the rating-quality contrast that prospective Phase 2 raters use to calibrate. 1 diff annotation, 1 peer annotation, rating-quality contrast.
Latest corpus-level findings: per-dimension divergence rates across peer pairs, cross-question outlier consistency with named patterns that fired, library entries per dimension. Versioned against corpus state; experimental pending Phase 2 validation.
Per-provider F1 from the latest calibration run, with visible sample size and stability caveat. Citable dataset; immutable per run.
Every Frame Check citation form in one place: the tool, the frame library, individual FVS entries, the methodology paper, and calibration runs. Plain text and BibTeX.
Applied analyses of specific public documents. 6 published entries.