Project MSC corrected complete-layer evidence

Technical summary

The corrected intervention zeros both energy and mask. Corrected validation selected M11 for all seven layers, so the frozen routed system equals M11. Official uniform-macro rMAE is 0.192936 for M11, 0.240423 for M10, and 0.214748 for M12-C.

Source: corrected_evaluation/metrics/official_evaluation_ensemble.csvTable: corrected_evaluation/metrics/official_evaluation_ensemble.csv

Corrected frozen-checkpoint evaluation.

Broad robustness

The corrected policy preserves M11 across ordinary and mixed states.

rMAE across broad corruption scenarios

Source: tables/ensemble_broad_scenarios.csvTable: tables/ensemble_broad_scenarios.csv

Read the exact packaged CSV rows used by this chart.

SELECT system, scenario, rmae, bias, r68, large_error_rate_gt20pct FROM read_csv_auto('tables/ensemble_broad_scenarios.csv')

rMAE across broad corruption scenarios data
Scenario	rMAE	System
Clean	0.18	M10
Random dead 5%	0.16	M10
Block dead 5%	0.18	M10
Layer gain drift	0.18	M10
Gaussian noise	0.19	M10
Hot cells	0.18	M10
Mixed unseen	0.19	M10
Clean	0.15	M11
Random dead 5%	0.16	M11
Block dead 5%	0.16	M11
Layer gain drift	0.15	M11
Gaussian noise	0.17	M11
Hot cells	0.15	M11
Mixed unseen	0.18	M11
Clean	0.18	M12-C
Random dead 5%	0.19	M12-C
Block dead 5%	0.19	M12-C
Layer gain drift	0.18	M12-C
Gaussian noise	0.18	M12-C
Hot cells	0.18	M12-C
Mixed unseen	0.2	M12-C
Clean	0.15	Routed
Random dead 5%	0.16	Routed
Block dead 5%	0.16	Routed
Layer gain drift	0.15	Routed
Gaussian noise	0.17	Routed
Hot cells	0.15	Routed
Mixed unseen	0.18	Routed

Physical complete-layer failures

All 120,800 events are evaluated once per layer with energy and mask entries zeroed.

Corrected complete-layer rMAE by failed layer

Source: corrected_evaluation/metrics/official_evaluation_ensemble.csvTable: corrected_evaluation/metrics/official_evaluation_ensemble.csv

Read the exact packaged CSV rows used by this chart.

SELECT system, layer_original_id AS layer, rmae, bias, r68, large_error_rate_gt20pct FROM read_csv_auto('corrected_evaluation/metrics/official_evaluation_ensemble.csv')

Corrected complete-layer rMAE by failed layer data
Failed layer	rMAE	System
0	0.19	M10
0	0.18	M11
0	0.19	M12-C
0	0.18	Routed
1	0.18	M10
1	0.16	M11
1	0.19	M12-C
1	0.16	Routed
2	0.32	M10
2	0.24	M11
2	0.26	M12-C
2	0.24	Routed
3	0.18	M10
3	0.17	M11
3	0.19	M12-C
3	0.17	Routed
12	0.32	M10
12	0.23	M11
12	0.25	M12-C
12	0.23	Routed
13	0.31	M10
13	0.21	M11
13	0.24	M12-C
13	0.21	Routed
14	0.18	M10
14	0.16	M11
14	0.18	M12-C
14	0.16	Routed

Corrected complete-layer aggregates

Source: tables/corrected_complete_layer_aggregates.csvTable: tables/corrected_complete_layer_aggregates.csv

Read the exact packaged CSV rows used by this table.

SELECT system, aggregate, rmae, bias, r68, large_error_rate_gt20pct FROM read_csv_auto('tables/corrected_complete_layer_aggregates.csv')

Corrected complete-layer aggregates
System	Average	rMAE	Bias	R68	Large-error
M10	uniform macro	0.24	0.02	0.27	0.44
M11	uniform macro	0.19	0.04	0.22	0.35
M12-C	uniform macro	0.21	-0.05	0.25	0.41
Routed	uniform macro	0.19	0.04	0.22	0.35
M10	voxel count weighted	0.28	-0	0.33	0.52
M11	voxel count weighted	0.21	0.05	0.23	0.38
M12-C	voxel count weighted	0.23	-0.04	0.27	0.45
Routed	voxel count weighted	0.21	0.05	0.23	0.38

Frozen validation decision

The expert choice was frozen from corrected validation predictions before official scoring.

Corrected validation selection

Source: tables/corrected_validation_ensemble.csvTable: tables/corrected_validation_ensemble.csv

Read the exact packaged CSV rows used by this table.

SELECT layer_original_id AS layer, system, rmae FROM read_csv_auto('tables/corrected_validation_ensemble.csv') WHERE system IN ('M11','M12_C_lambda_010')

Corrected validation selection
Layer	Candidate	Validation rMAE
0	M11	0.18
0	M12_C_lambda_010	0.19
1	M11	0.16
1	M12_C_lambda_010	0.19
2	M11	0.24
2	M12_C_lambda_010	0.26
3	M11	0.16
3	M12_C_lambda_010	0.18
12	M11	0.23
12	M12_C_lambda_010	0.25
13	M11	0.21
13	M12_C_lambda_010	0.23
14	M11	0.16
14	M12_C_lambda_010	0.18

Detector-status inconsistency control

Historical mask-only rows retain clean layer energies and are not complete-layer evidence.

Physical failure and mask-only inconsistency

Source: tables/corrected_vs_mask_only_aggregates.csvTable: tables/corrected_vs_mask_only_aggregates.csv

Read the exact packaged CSV rows used by this chart.

SELECT system, aggregate, corrected_rmae, mask_only_rmae FROM read_csv_auto('tables/corrected_vs_mask_only_aggregates.csv') WHERE aggregate = 'uniform_macro'

Physical failure and mask-only inconsistency data
System	Uniform macro rMAE	Intervention
M10	0.24	Corrected energy + mask
M10	0.83	Mask-only inconsistency
M11	0.19	Corrected energy + mask
M11	0.79	Mask-only inconsistency
M12-C	0.21	Corrected energy + mask
M12-C	0.2	Mask-only inconsistency

Limitations and next steps

The data are simulated, masks are assumed accurate, M12-C was originally selected using the historical mask-only criterion, and no mask/no-mask ablation exists. The result supports broad M11 robustness, not universal superiority.

Further questions

Would corrected-target selection or retraining improve M12-C? How robust is M11 to mask errors and real detector data?

Sources

corrected_evaluation/metrics/official_evaluation_ensemble.csvcorrected_evaluation/metrics/official_evaluation_ensemble.csv
Corrected frozen-checkpoint evaluation.
tables/ensemble_broad_scenarios.csvtables/ensemble_broad_scenarios.csv
Frozen broad-corruption evaluation.
tables/corrected_vs_mask_only_aggregates.csvtables/corrected_vs_mask_only_aggregates.csv
Physical failure versus inconsistent metadata control.