Component Logit Attribution

The decomposition reconstructs the model logits

For each input, the report computes residual + head 0 + head 1 + head 2 + head 3 at the answer position and compares it with the model's actual logits.

Validation

Max reconstruction error: 3.41061e-13

Sampled accuracy: 1.0000

Mean final margin: 15.79

Average component contributions — Average contribution to the final winning margin on the stratified sample.

Contribution by max digit

The attribution uses the model's final strongest wrong digit for each input, so every component's margin contribution adds exactly to the model's final margin.

Diagnostic examples

Diagnostic stacked margin attribution — Bars show the additive components; black dots show the reconstructed total margin.

Input	Component	Correct logit	Own top	Own margin	Final-margin contribution
`[3, 7, 2, 5, 1]`	residual	2.31	7	0.46	0.66
`[3, 7, 2, 5, 1]`	head 0	31.37	3	-23.43	9.36
`[3, 7, 2, 5, 1]`	head 1	15.17	3	-0.72	3.63
`[3, 7, 2, 5, 1]`	head 2	3.78	9	-11.23	6.10
`[3, 7, 2, 5, 1]`	head 3	35.31	6	-23.38	-2.89
`[3, 7, 2, 5, 1]`	total	87.94	7	16.86	16.86
`[9, 1, 2, 3, 4]`	residual	0.98	7	-1.33	-0.87
`[9, 1, 2, 3, 4]`	head 0	20.80	9	12.73	19.02
`[9, 1, 2, 3, 4]`	head 1	-4.29	3	-20.18	-10.55
`[9, 1, 2, 3, 4]`	head 2	35.21	9	12.97	12.97
`[9, 1, 2, 3, 4]`	head 3	100.64	6	-11.30	-2.74
`[9, 1, 2, 3, 4]`	total	153.35	9	17.83	17.83
`[5, 4, 5, 1, 0]`	residual	1.65	7	-0.66	0.26
`[5, 4, 5, 1, 0]`	head 0	22.58	3	-33.65	37.88
`[5, 4, 5, 1, 0]`	head 1	11.54	3	-4.34	9.66
`[5, 4, 5, 1, 0]`	head 2	9.62	2	-59.59	25.07
`[5, 4, 5, 1, 0]`	head 3	107.78	6	-57.78	-57.78
`[5, 4, 5, 1, 0]`	total	153.16	5	15.09	15.09
`[0, 0, 0, 0, 0]`	residual	-2.15	7	-4.46	-1.16
`[0, 0, 0, 0, 0]`	head 0	17.12	3	-39.11	-27.87
`[0, 0, 0, 0, 0]`	head 1	-3.98	3	-19.87	-10.44
`[0, 0, 0, 0, 0]`	head 2	51.80	2	-17.46	-16.50
`[0, 0, 0, 0, 0]`	head 3	300.11	0	68.38	68.38
`[0, 0, 0, 0, 0]`	total	362.89	0	12.40	12.40
`[0, 1, 0, 0, 0]`	residual	-0.99	7	-3.30	1.16
`[0, 1, 0, 0, 0]`	head 0	44.99	3	-11.23	27.87
`[0, 1, 0, 0, 0]`	head 1	6.46	3	-9.43	10.44
`[0, 1, 0, 0, 0]`	head 2	68.30	2	-0.96	16.50
`[0, 1, 0, 0, 0]`	head 3	154.30	0	-45.53	-45.53
`[0, 1, 0, 0, 0]`	total	273.06	1	10.46	10.46

Artifacts

Diagnostic components CSV Sample contributions CSV Aggregate CSV Metadata JSON