File size: 728 Bytes
c8beaf3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 | # Aggregation replay protocol
Hold source rows, row embeddings, query embeddings, taxonomy, query roster and
similarity measure fixed. Change only the mapping from row embeddings to ranked
taxon scores.
At minimum, declare whether the method uses one species centroid, instance-max,
source-separated prototypes, multiple learned or clustered prototypes, or
another rule. Report prototype count, memory footprint, tie-breaking and any
hyperparameters. Method selection must not use the canonical QT26-QC test labels.
Report the same micro, macro, top-k and transition endpoints as gallery-state
replay. A method chosen after inspecting Fish-Vista is diagnostic on that roster,
not an independent museum-domain holdout result.
|