## Token classes (mean): V-measure / many-to-one / sense accuracy
tier 3000
JT-V2D 0.624 / 0.641 / 0.563
JT-V2D-COLL 0.665 / 0.659 / 0.592
JT-V3 0.620 / 0.641 / 0.571
JT-V3-COLL 0.662 / 0.660 / 0.591
JT-V2D-SHUF 0.370 / 0.471 / 0.329
JT-V2D-COLL-SHUF 0.464 / 0.508 / 0.375
JT-V3-SHUF 0.367 / 0.468 / 0.327
JT-V3-COLL-SHUF 0.462 / 0.504 / 0.366
MAJ 0.975 / 0.987 / 0.875
tier 10000
JT-V2D 0.806 / 0.786 / 0.631
JT-V2D-COLL 0.825 / 0.797 / 0.661
JT-V3 0.787 / 0.764 / 0.654
JT-V3-COLL 0.827 / 0.789 / 0.689
JT-V2D-SHUF 0.466 / 0.551 / 0.395
JT-V2D-COLL-SHUF 0.558 / 0.593 / 0.454
JT-V3-SHUF 0.459 / 0.542 / 0.364
JT-V3-COLL-SHUF 0.554 / 0.584 / 0.448
MAJ 0.977 / 0.988 / 0.885
tier 30000
JT-V2D 0.864 / 0.835 / 0.643
JT-V2D-COLL 0.869 / 0.837 / 0.630
JT-V3 0.826 / 0.789 / 0.652
JT-V3-COLL 0.870 / 0.810 / 0.661
JT-V2D-SHUF 0.614 / 0.671 / 0.490
JT-V2D-COLL-SHUF 0.687 / 0.709 / 0.529
JT-V3-SHUF 0.577 / 0.633 / 0.448
JT-V3-COLL-SHUF 0.658 / 0.673 / 0.503
MAJ 0.977 / 0.988 / 0.886
## Split diagnostics (mean over seeds): eligible, split, split precision, split recall, base rate
3000 JT-V3 159.2 0.2 0.000 0.000 0.102
3000 JT-V3-SHUF 159.2 0.2 0.000 0.000 0.102
10000 JT-V3 504.8 5.6 0.033 0.003 0.130
10000 JT-V3-SHUF 504.8 4.0 0.029 0.003 0.130
30000 JT-V3 1124.6 17.4 0.112 0.013 0.133
30000 JT-V3-SHUF 1124.6 25.2 0.130 0.021 0.133
## Parse link F1 (mean)
3000: ATTN-V2D 0.516 ATTN-V3 0.515
10000: ATTN-V2D 0.609 ATTN-V3 0.604
30000: ATTN-V2D 0.642 ATTN-V3 0.632
## Predictions (at 30000)
S1: FAIL — split precision 0.112 (bar 0.266 = 2 x base), recall 0.013 (bar 0.3), words split 17.4 of 1124.6
S2: FAIL — collapsed sense V3 − V2D +0.031 [-0.060, +0.122]
S3: FAIL — sense V3 − V2D +0.009 [-0.110, +0.128]
S4: FAIL — words split shuffled 25.2 vs ordered 17.4 (bar <= 0.5x); shuffled sense V3 − V2D -0.042 [-0.119, +0.035]
S5: FAIL — V-measure V3 − V2D -0.038 [-0.060, -0.015]
S6: PASS — 60/60 audit checks PASS