Our prior report (Technical Report 1) found that unbounded ledger-based replay strongly preserves early-task accuracy in a sequential learning setting, at the cost of slower acquisition of the newest task — a measured stability-plasticity trade-off, observed on synthetic tabular data. This report tests whether the same finding holds on real image data. Using a small feedforward network on real 8×8 grayscale handwritten digit images from a standard, publicly available dataset, we repeated the identical three-condition experiment (no replay, bounded 50-example buffer, unbounded ledger replay) across four sequential binary digit-pair classification tasks. Unbounded replay again dominated on retention (100.0%, 98.6%, 98.6%, 100.0% across the four tasks versus 65.3–100.0% and 72.2–100.0% for the two baselines), but — unlike the synthetic-data experiment — showed no measurable plasticity cost on the newest task. We report and discuss this divergence honestly rather than treating either result as the final word.
A single positive result on synthetic data is weak evidence. Real image data introduces higher input dimensionality (64 pixel features versus 7 synthetic features), different class separability, and a more realistic test of whether the replay mechanism's benefits are an artifact of the synthetic task design. This report repeats the Technical Report 1 protocol on real images with no other changes to the experimental design beyond the input data source and the number of training epochs, which was increased from a fixed step budget to three passes over each task's training data to accommodate the smaller per-task sample count in the image dataset.
We used a small, publicly available dataset of 8×8 grayscale handwritten digit images (1,797 samples total, digits 0–9), flattened to 64-dimensional pixel vectors and normalized to [0,1]. Four sequential binary tasks were constructed, each distinguishing one pair of digits: (0,1), (2,3), (4,5), (6,7). An 80/20 train/test split was applied within each task. The same three replay conditions from Technical Report 1 were re-implemented for this input size (a 64-32-1 feedforward network) with the same replay mechanics: 3 replayed examples per training step for the bounded and unbounded conditions, differing only in whether the replay pool evicts old examples (bounded, capacity 50) or retains all history (unbounded/ledger).
| Condition | Task 1 | Task 2 | Task 3 | Task 4 |
|---|---|---|---|---|
| No replay | 65.3% | 70.8% | 97.3% | 100.0% |
| Bounded RAM (n=50) | 72.2% | 58.3% | 97.3% | 100.0% |
| Unbounded ledger | 100.0% | 98.6% | 98.6% | 100.0% |
Table 1 — Final test accuracy per real-image task after sequential training on all four tasks.
Unbounded ledger replay reached 98.6–100.0% across all four tasks, including the newest one — dominating both baselines on every task simultaneously. This contrasts with Technical Report 1, where the unbounded condition traded away newest-task accuracy (53.3%) in exchange for old-task retention.
We do not have a confirmed causal explanation, and we are explicit about that. Two differences between this experiment and Technical Report 1 are plausible contributors, and distinguishing between them is left to future work:
We consider it likely that the plasticity cost identified in Technical Report 1 is a real phenomenon that depends on the ratio of new-task training steps to accumulated replay volume, rather than a fixed, universal property of unbounded replay. This report does not isolate that ratio as a controlled variable; doing so is the natural next experiment.
The core retention advantage of unbounded ledger replay — found on synthetic data in Technical Report 1 — replicated on real image data. The plasticity cost observed previously did not replicate under this experimental configuration. Both results are reported as observed; we recommend a controlled follow-up varying training-steps-per-task and task difficulty independently before drawing a general conclusion about when the trade-off appears.