# E114 continual-learning report

Five-task streams (D=20), sealed seeds 101–110. All models were trained here with numpy on CPU.

| learner | final acc | forgetting | BWT | zero-shot | train sample-steps | memory bytes |
|---|---|---|---|---|---|---|
| experts | 0.919 | 0.000 | 0.000 | 0.573 | 90000 | 26880 |
| knn_memory | 0.798 | -0.000 | 0.000 | 0.537 | 0 | 252000 |
| naive | 0.684 | 0.286 | -0.286 | 0.583 | 90000 | 0 |
| replay | 0.823 | 0.097 | -0.097 | 0.580 | 126000 | 50400 |

Outer loop: `max final accuracy s.t. mean forgetting <= 0.05; tie -> less CPU` chose **experts** on dev (eligible ['experts', 'knn_memory']);
sealed confirmation 10/10, p = 0.00195.
40 training records in `raw/model_registry.json`. kNN memory costs 7800000 distance operations per stream at inference.
Limits: no forward transfer beyond chance-level zero-shot; synthetic tasks; no language-model fine-tuning (no GPU, NeMo BLOCKED).
