# Corrected deterministic complete-layer failure evaluation

This evaluation sets both every voxel energy and every operational-mask entry
of the selected layer to zero. The historical intervention is preserved and
relabelled **detector-status inconsistency / mask-only intervention**.

Large error is defined as `abs((prediction - truth) / truth) > 0.20`.
The per-layer M11/M12-C policy was selected from 18,120 validation
events and frozen at SHA-256 `8fb737e07f75471b4fe602fe358e0da34e2b3ace9a7ffc67002135a7a3d18f80` before the 120,800
official-evaluation targets were loaded.

## Frozen route

```json
{
  "0": "M11",
  "1": "M11",
  "12": "M11",
  "13": "M11",
  "14": "M11",
  "2": "M11",
  "3": "M11"
}
```

## Official-evaluation ensemble results

| system | layer | rmae | bias | r68 | large_error_rate |
| --- | --- | --- | --- | --- | --- |
| M10 | 0 | 0.194574 | 0.072709 | 0.197543 | 0.362343 |
| M11 | 0 | 0.179279 | 0.051319 | 0.209010 | 0.333079 |
| M12_C_lambda_010 | 0 | 0.194182 | -0.047656 | 0.224834 | 0.370108 |
| routed_M11_M12_C | 0 | 0.179279 | 0.051319 | 0.209010 | 0.333079 |
| M10 | 1 | 0.182474 | 0.028213 | 0.205096 | 0.335116 |
| M11 | 1 | 0.164105 | 0.032673 | 0.177747 | 0.281291 |
| M12_C_lambda_010 | 1 | 0.191663 | -0.054827 | 0.213952 | 0.357202 |
| routed_M11_M12_C | 1 | 0.164105 | 0.032673 | 0.177747 | 0.281291 |
| M10 | 2 | 0.323840 | -0.109672 | 0.364518 | 0.616929 |
| M11 | 2 | 0.239318 | 0.048199 | 0.276678 | 0.425472 |
| M12_C_lambda_010 | 2 | 0.260270 | -0.064280 | 0.295613 | 0.500257 |
| routed_M11_M12_C | 2 | 0.239318 | 0.048199 | 0.276678 | 0.425472 |
| M10 | 3 | 0.178383 | 0.058498 | 0.188121 | 0.306581 |
| M11 | 3 | 0.166267 | 0.028512 | 0.189768 | 0.301035 |
| M12_C_lambda_010 | 3 | 0.186290 | -0.047600 | 0.211338 | 0.351217 |
| routed_M11_M12_C | 3 | 0.166267 | 0.028512 | 0.189768 | 0.301035 |
| M10 | 12 | 0.319303 | -0.030685 | 0.382343 | 0.583055 |
| M11 | 12 | 0.228957 | 0.059541 | 0.259993 | 0.437922 |
| M12_C_lambda_010 | 12 | 0.251727 | -0.037802 | 0.298920 | 0.488518 |
| routed_M11_M12_C | 12 | 0.228957 | 0.059541 | 0.259993 | 0.437922 |
| M10 | 13 | 0.305257 | 0.060742 | 0.370690 | 0.537012 |
| M11 | 13 | 0.209746 | 0.050581 | 0.226053 | 0.383063 |
| M12_C_lambda_010 | 13 | 0.236213 | -0.020329 | 0.270658 | 0.442483 |
| routed_M11_M12_C | 13 | 0.209746 | 0.050581 | 0.226053 | 0.383063 |
| M10 | 14 | 0.179133 | 0.073882 | 0.179832 | 0.310621 |
| M11 | 14 | 0.162880 | 0.036960 | 0.180549 | 0.284934 |
| M12_C_lambda_010 | 14 | 0.182893 | -0.045636 | 0.208817 | 0.341623 |
| routed_M11_M12_C | 14 | 0.162880 | 0.036960 | 0.180549 | 0.284934 |

## Complete-layer aggregate ensemble results

| system | aggregate | rmae | bias | r68 | large_error_rate |
| --- | --- | --- | --- | --- | --- |
| M10 | uniform_macro | 0.240423 | 0.021955 | 0.269735 | 0.435951 |
| M10 | voxel_count_weighted | 0.284442 | -0.002658 | 0.333851 | 0.518053 |
| M11 | uniform_macro | 0.192936 | 0.043969 | 0.217114 | 0.349542 |
| M11 | voxel_count_weighted | 0.210393 | 0.048844 | 0.234590 | 0.384004 |
| M12_C_lambda_010 | uniform_macro | 0.214748 | -0.045447 | 0.246305 | 0.407344 |
| M12_C_lambda_010 | voxel_count_weighted | 0.234635 | -0.041105 | 0.270250 | 0.446443 |
| routed_M11_M12_C | uniform_macro | 0.192936 | 0.043969 | 0.217114 | 0.349542 |
| routed_M11_M12_C | voxel_count_weighted | 0.210393 | 0.048844 | 0.234590 | 0.384004 |

Full-precision per-seed, ensemble, aggregate, prediction, and corrected-versus-
mask-only files are stored in this directory.

## Individual-seed aggregate mean ± sample standard deviation

| System | Aggregate | Seeds | rMAE | Bias | R68 | Large-error rate |
| --- | --- | ---: | ---: | ---: | ---: | ---: |
| M10 | uniform_macro | 3 | 0.250586 ± 0.006165 | 0.021955 ± 0.023231 | 0.281001 ± 0.004647 | 0.454558 ± 0.004746 |
| M10 | voxel_count_weighted | 3 | 0.295554 ± 0.009993 | -0.002658 ± 0.037018 | 0.342302 ± 0.009385 | 0.534047 ± 0.003999 |
| M11 | uniform_macro | 5 | 0.209026 ± 0.009635 | 0.043969 ± 0.034609 | 0.232664 ± 0.005449 | 0.384548 ± 0.019066 |
| M11 | voxel_count_weighted | 5 | 0.226501 ± 0.011625 | 0.048844 ± 0.036762 | 0.252014 ± 0.008051 | 0.415605 ± 0.018578 |
| M12_C_lambda_010 | uniform_macro | 5 | 0.234588 ± 0.007655 | -0.045447 ± 0.020856 | 0.270319 ± 0.009410 | 0.455817 ± 0.015153 |
| M12_C_lambda_010 | voxel_count_weighted | 5 | 0.254147 ± 0.008656 | -0.041105 ± 0.021630 | 0.292662 ± 0.010169 | 0.489222 ± 0.014827 |
| routed_M11_M12_C | uniform_macro | 5 | 0.209026 ± 0.009635 | 0.043969 ± 0.034609 | 0.232664 ± 0.005449 | 0.384548 ± 0.019066 |
| routed_M11_M12_C | voxel_count_weighted | 5 | 0.226501 ± 0.011625 | 0.048844 ± 0.036762 | 0.252014 ± 0.008051 | 0.415605 ± 0.018578 |

Ensemble metrics above are calculated after averaging event-level predictions and are not averages of these individual-seed metrics.

## Corrected versus historical status inconsistency

The historical arrays are preserved under the label **detector-status inconsistency / mask-only intervention**. Full-precision per-layer and aggregate comparisons are in `metrics/corrected_vs_mask_only_ensemble.csv` and `metrics/corrected_vs_mask_only_aggregates.csv`. For M10 and M11, contradictory mask-only metadata is substantially more damaging than physically zeroing the same layer; for M12-C it produces slightly lower rMAE but does not represent missing measurements.