- Match %
- Jev's model-output score (0–100%) for "does this window match the declared episode?" Higher means the model sees stronger similarity. This is not a probability of any market outcome.
- Recall@3
- Of the top 3 ranked candidates, how many are true matches? For example, "2 / 3" means 2 of the top 3 results were genuine episodes of that type. Maximum equals the total true matches (up to 3).
- MRR (Mean Reciprocal Rank)
- Where the first true match appears in the ranked list. 1.00 = first result is a match. 0.50 = first match is at position 2. 0.33 = position 3. Measures how quickly you find a relevant result.
- Brier Score
- Average squared error between the model's match score and the true label (1 for match, 0 for distractor). Lower is better. Under 0.15 is strong; over 0.30 suggests poor calibration.
- True matches in corpus
- How many of the 40 synthetic windows are seeded as that episode type. This is the ground truth count the ranking is evaluated against.
- Source tag (jev / deterministic / unavailable)
- "jev" = ranked by the Jev model, a real match score. "deterministic" = the deterministic baseline ranking used only when Jev is unavailable for every candidate; its score is derived from rank position, not a match probability. "unavailable" = Jev did not return a usable score for this specific candidate.