Datasets:
Final locked standard-data results
All 260 predeclared runs completed: thirteen methods, ten paired fresh seeds, and two datasets. Each stream has five two-class arrivals with a shared ten-way output and no task ID at inference. The source lock, official-test predictions, pre/post-acquisition checkpoints, and all refusal records are retained.
Ordinary baselines fit all 54,000 nonvalidation images; certificate methods spend 6,000 of that observed-data allowance on independent calibration and fit 48,000. The center ablation uses the same 48,000 fitting images. All continual learners obey a 6 MiB persistent payload budget; the explicitly noncontinual joint reference retains all seen data. These are CPU studies on MNIST/Fashion-MNIST, not large-scale validation or a state-of-the-art claim.
MNIST
Accuracy is final official-test micro accuracy. Intervals are descriptive Student-t 95% intervals across ten seed replications.
| Method | Mean accuracy | SD | 95% interval | Mean protected classes |
|---|---|---|---|---|
| SGD | 19.312% | 1.440 pp | [18.282, 20.342]% | 0.0 |
| ER | 89.965% | 1.156 pp | [89.138, 90.792]% | 0.0 |
| DER++ | 92.028% | 0.977 pp | [91.329, 92.727]% | 0.0 |
| ER-ACE, 2 epochs | 92.669% | 1.623 pp | [91.508, 93.830]% | 0.0 |
| ER-ACE, 4 epochs | 95.155% | 0.380 pp | [94.883, 95.427]% | 0.0 |
| Cumulative RFF ridge | 89.443% | 0.296 pp | [89.231, 89.655]% | 0.0 |
| Global prototypes | 96.476% | 0.121 pp | [96.389, 96.563]% | 0.0 |
| Scalar center, risk 0.1 | 19.932% | 0.892 pp | [19.294, 20.570]% | 10.0 |
| Center, controller disabled | 81.177% | 1.430 pp | [80.154, 82.200]% | 0.0 |
| Immutable regions, risk 0.1 | 36.609% | 6.796 pp | [31.747, 41.471]% | 2.6 |
| Fresh renewal, risk 0.2 | 92.720% | 7.874 pp | [87.087, 98.353]% | 9.6 |
| Fresh renewal, risk 0.1 | 46.548% | 22.557 pp | [30.412, 62.684]% | 4.6 |
| Stagewise joint reference | 93.119% | 0.393 pp | [92.838, 93.400]% | 0.0 |
Protected-class counts are certificate-admission counts, not the number of accurately learned classes. The scalar-center certificate concerns preservation of acquired center predictions, including incorrect ones. Ordinary controls have no certificate and are assigned zero in this column.
| Certificate policy | Difference from ER-ACE, 2 epochs | Difference from global prototypes |
|---|---|---|
| Scalar center, risk 0.1 | -72.737 pp [-73.750, -71.724] | -76.544 pp [-77.172, -75.916] |
| Immutable regions, risk 0.1 | -56.060 pp [-61.240, -50.880] | -59.867 pp [-64.768, -54.966] |
| Fresh renewal, risk 0.2 | +0.051 pp [-5.789, +5.891] | -3.756 pp [-9.358, +1.846] |
| Fresh renewal, risk 0.1 | -46.121 pp [-62.152, -30.090] | -49.928 pp [-66.057, -33.799] |
Intervals use per-seed paired differences; they are not multiplicity-adjusted tests of superiority. The four-epoch ER-ACE control is explicitly compute-relaxed.
| Method | Mean fitting time, s | Mean calibration time, s | Serial batch-one latency, ms | Mean peak persistent payload, MiB |
|---|---|---|---|---|
| SGD | 1.219 | 0.000 | 0.051 | 0.452 |
| ER | 2.694 | 0.000 | 0.056 | 5.997 |
| DER++ | 4.050 | 0.000 | 0.051 | 5.997 |
| ER-ACE, 2 epochs | 2.904 | 0.000 | 0.073 | 5.997 |
| ER-ACE, 4 epochs | 5.960 | 0.000 | 0.072 | 5.997 |
| Cumulative RFF ridge | 2.517 | 0.000 | 0.128 | 5.139 |
| Global prototypes | 7.130 | 0.000 | 4.996 | 3.846 |
| Scalar center, risk 0.1 | 4.590 | 0.313 | 0.249 | 3.961 |
| Center, controller disabled | 5.547 | 0.000 | 0.279 | 4.466 |
| Immutable regions, risk 0.1 | 9.607 | 0.217 | 0.477 | 1.461 |
| Fresh renewal, risk 0.2 | 7.085 | 0.348 | 4.367 | 5.975 |
| Fresh renewal, risk 0.1 | 7.063 | 0.339 | 0.461 | 5.975 |
| Stagewise joint reference | 3.838 | 0.000 | 0.051 | 41.651 |
Fitting/calibration wall times may reflect four-worker contention. Serial latency was measured after training workers exited, using fixed final checkpoints, three warmups, and twenty-one trials. Persistent payload excludes transient buffers and Python overhead; per-process peak RSS is separately retained. Joint-reference payload includes its retained raw data and is exempt from the continual-memory budget.
Fresh renewal, risk 0.2 accepted 48/50 proposed checkpoints, including the final ten-class proposal in 8/10 runs. Its own unvalidated full-proposal final accuracy averaged 96.428%; this diagnostic uses identical fitted prototypes and isolates the acceptance controller from the ordinary baseline's additional fitting data.
Fresh renewal, risk 0.1 accepted 19/50 proposed checkpoints, including the final ten-class proposal in 0/10 runs. Its own unvalidated full-proposal final accuracy averaged 96.428%; this diagnostic uses identical fitted prototypes and isolates the acceptance controller from the ordinary baseline's additional fitting data.
FASHION MNIST
Accuracy is final official-test micro accuracy. Intervals are descriptive Student-t 95% intervals across ten seed replications.
| Method | Mean accuracy | SD | 95% interval | Mean protected classes |
|---|---|---|---|---|
| SGD | 18.983% | 3.061 pp | [16.793, 21.173]% | 0.0 |
| ER | 78.287% | 1.428 pp | [77.266, 79.308]% | 0.0 |
| DER++ | 78.628% | 1.867 pp | [77.292, 79.964]% | 0.0 |
| ER-ACE, 2 epochs | 80.012% | 1.822 pp | [78.709, 81.315]% | 0.0 |
| ER-ACE, 4 epochs | 82.597% | 1.046 pp | [81.849, 83.345]% | 0.0 |
| Cumulative RFF ridge | 80.123% | 0.209 pp | [79.973, 80.273]% | 0.0 |
| Global prototypes | 84.635% | 0.177 pp | [84.508, 84.762]% | 0.0 |
| Scalar center, risk 0.1 | 19.455% | 0.981 pp | [18.753, 20.157]% | 10.0 |
| Center, controller disabled | 66.187% | 2.109 pp | [64.678, 67.696]% | 0.0 |
| Immutable regions, risk 0.1 | 29.297% | 5.721 pp | [25.204, 33.390]% | 2.2 |
| Fresh renewal, risk 0.2 | 35.503% | 21.542 pp | [20.093, 50.913]% | 3.6 |
| Fresh renewal, risk 0.1 | 17.918% | 4.177 pp | [14.930, 20.906]% | 1.6 |
| Stagewise joint reference | 81.145% | 1.906 pp | [79.782, 82.508]% | 0.0 |
Protected-class counts are certificate-admission counts, not the number of accurately learned classes. The scalar-center certificate concerns preservation of acquired center predictions, including incorrect ones. Ordinary controls have no certificate and are assigned zero in this column.
| Certificate policy | Difference from ER-ACE, 2 epochs | Difference from global prototypes |
|---|---|---|
| Scalar center, risk 0.1 | -60.557 pp [-62.081, -59.033] | -65.180 pp [-65.917, -64.443] |
| Immutable regions, risk 0.1 | -50.715 pp [-54.622, -46.808] | -55.338 pp [-59.394, -51.282] |
| Fresh renewal, risk 0.2 | -44.509 pp [-60.730, -28.288] | -49.132 pp [-64.588, -33.676] |
| Fresh renewal, risk 0.1 | -62.094 pp [-65.440, -58.748] | -66.717 pp [-69.771, -63.663] |
Intervals use per-seed paired differences; they are not multiplicity-adjusted tests of superiority. The four-epoch ER-ACE control is explicitly compute-relaxed.
| Method | Mean fitting time, s | Mean calibration time, s | Serial batch-one latency, ms | Mean peak persistent payload, MiB |
|---|---|---|---|---|
| SGD | 1.181 | 0.000 | 0.055 | 0.452 |
| ER | 2.719 | 0.000 | 0.052 | 5.997 |
| DER++ | 4.134 | 0.000 | 0.055 | 5.997 |
| ER-ACE, 2 epochs | 3.102 | 0.000 | 0.086 | 5.997 |
| ER-ACE, 4 epochs | 5.873 | 0.000 | 0.085 | 5.997 |
| Cumulative RFF ridge | 2.578 | 0.000 | 0.123 | 5.139 |
| Global prototypes | 7.093 | 0.000 | 5.103 | 3.846 |
| Scalar center, risk 0.1 | 4.436 | 0.321 | 0.272 | 3.961 |
| Center, controller disabled | 5.246 | 0.000 | 0.293 | 4.466 |
| Immutable regions, risk 0.1 | 9.106 | 0.214 | 0.437 | 1.307 |
| Fresh renewal, risk 0.2 | 6.775 | 0.350 | 0.383 | 5.975 |
| Fresh renewal, risk 0.1 | 6.825 | 0.346 | 0.209 | 5.975 |
| Stagewise joint reference | 3.924 | 0.000 | 0.051 | 41.651 |
Fitting/calibration wall times may reflect four-worker contention. Serial latency was measured after training workers exited, using fixed final checkpoints, three warmups, and twenty-one trials. Persistent payload excludes transient buffers and Python overhead; per-process peak RSS is separately retained. Joint-reference payload includes its retained raw data and is exempt from the continual-memory budget.
Fresh renewal, risk 0.2 accepted 18/50 proposed checkpoints, including the final ten-class proposal in 0/10 runs. Its own unvalidated full-proposal final accuracy averaged 84.530%; this diagnostic uses identical fitted prototypes and isolates the acceptance controller from the ordinary baseline's additional fitting data.
Fresh renewal, risk 0.1 accepted 8/50 proposed checkpoints, including the final ten-class proposal in 0/10 runs. Its own unvalidated full-proposal final accuracy averaged 84.530%; this diagnostic uses identical fitted prototypes and isolates the acceptance controller from the ordinary baseline's additional fitting data.
Scope of the evidence
The guarantee allocation is 0.05 per seeded run and per policy across all its paid class tests, including failed attempts. It is not a joint 95% assertion across all datasets, seeds, and policies. Renewal certifies each accepted checkpoint under its independence/distribution assumptions; it does not certify unaccepted proposals, their fitting iterates, or ever-error on a fixed old example. Unused acquired calibration archives and all retained proposed prototypes are included in the memory count.
Full class/task matrices, macro accuracy, signed forgetting, backward/forward transfer, stagewise joint intransigence, ECE15, source IDs, confidence ledgers, memory payload, and arithmetic component counts are in the JSON/NPZ captures. Arithmetic counts are component estimates rather than total hardware FLOPs. NLL/Brier are a separately hashed post-lock descriptive analysis of saved predictions, without temperature fitting or model-selection feedback.
No final outcomes changed the source, method list, learning rates, feature maps, prototype counts, risk settings, calibration allocation, or seed list. All earlier development failures remain preserved.