Full prepared pipeline input. The default filter shows only member-resolved rows that received a JPAR update in this cutoff run.

Click a person row to overlay their trajectory and score position across the diagnostics; click it again to clear the selection.

Lines show p10, p25, median, p75, and p90 of each system's as-of score distribution after each event. Median is black.

Histograms use the final person-level scores for each system, clipped to the 1st and 99th percentiles to keep long tails readable.

Each scatter plot compares original fiducial JPAR rank on the x-axis against another system's rank on the y-axis. Lower rank is better, so closer to the upper-left is stronger.

JPAR's adjustment factor is event mean time / anchor-implied mean time. Below 1.00 inflates every participant's raw event JPAR (a relative penalty); above 1.00 lowers it (a relative boost). These are calibration shifts, not a claim that a puzzle itself was objectively easy or hard.

Median JPAR delta = adjusted_event_jpar - raw_event_jpar: a positive value increases JPAR scores (worse); a negative value decreases them (better).

Negative final calibration effect means JPAR ended lower (better) than the same running update using raw event JPAR; positive means it ended higher (worse). Click headers to sort.

Raw event JPAR is computed from the event mean; adjusted event JPAR applies JPAR's calibration multiplier. The divergence between their medians is the calibration intervention for that event. Green shows the as-of median latest JPAR after each event.