Prototype status
Loading the saved R4 prototype data…
Read-only record inspection
Inspect a topic and saved ordering
Start with Baseline, then choose a saved configuration. Filters help you inspect records; they preserve the saved ranks and do not change the metrics.
View the saved configuration and metric conventions
The values above describe the saved unfiltered top 10. The complete saved ranking contains all 200 candidates. No metric or score is recalculated in this viewer.
A saved rank can remain visible after earlier records are filtered out. Filtering is an inspection aid, not a new ranked result list.
Original frozen comparison · All 50 topics
Reading the saved comparison
The original Full and Guardrail configurations have lower mean relevance than Baseline. Their paired intervals include zero and do not establish equivalence or validate adoption. Baseline remains the default.
The cards show saved mean nDCG@10, a relevance measure that rewards placing more relevant documents near the top. Higher values indicate greater measured relevance under the retained judgments.
The two text-only development profiles are available in the selector for inspection. All 50 topics were used during development; their saved results do not provide independent confirmation and no profile is promoted.
Candidate coverage and judgment limits
Counts describe candidate appearances. A document can appear under several topics and have different task-specific judgments.
TREC-COVID human relevance grades are 0, 1 and 2. Unjudged remains null, visibly distinct from a human grade 0. For saved nDCG@10, unjudged records receive zero gain for scoring only; the ideal denominator comes from the matching full-collection judgments.
Corpus text is from July 2020; retained OpenAlex metadata is from October 2026. An OA flag is metadata, not an assurance of full-text access, institutional entitlement or accessibility. Filters stay in your browser and are not submitted to a service.