The /evaluation workbench runs a three-stage protocol on isolated CPU copies of a selected MicroModel:

1. Measure the starting model.
2. Train on the initial experience split; measure again.
3. Train on feedback; measure again.

The runner can update all parameters in its isolated copy or only a selected mounted weight module. It never promotes weights into the active model. Context drones, retrieval and general-model providers are excluded from this protocol so that parameter learning can be measured separately.

Evaluation contracts

• Retention: unchanged examples from the initial learning session. This measures preservation of learned responses, not held-out transfer.
• Feedback uptake: correct responses to the feedback prompts, scored with the revised targets both before and after correction.
• Transfer: prompts absent from both training sessions. In the bundled synthetic task these are unseen prompt forms for familiar mappings, not new domains or new semantic knowledge.
• Response loss: cross-entropy over response bytes and the terminating newline. Prompt tokens are excluded from the training objective.
• Exact accuracy: greedy generation receives the prompt only and stops at newline or its configured generation budget. Expected answers are used separately for evaluation loss and correctness checks.
• Retention change: post-feedback retention minus post-initial-learning retention, in percentage points. Negative values indicate forgetting.
• Feedback gain: post-feedback corrected-answer accuracy minus pre-feedback corrected-answer accuracy.

Reject duplicate split prompts, transfer/training overlap, conflicting retention targets, examples too long for the model context, and generation budgets too short for the expected response.

Experience replay

An optional rehearsal fraction mixes unchanged initial-training examples into the feedback session. Corrected prompts are excluded so that obsolete answers are not rehearsed. Transfer probes are never part of replay. This is a bounded rehearsal experiment using saved initial examples, not an implemented automatic memory-consolidation policy.

Both conditions use the same training-step and batch-size budget. Replay changes the distribution of example exposures; it does not add training steps.

First controlled comparison

See [private artifact]. Both conditions used identical starting weights and identical post-initial-learning weights for each of three seeds. The task has four initial signal/action mappings, two corrections, two unchanged retention probes and four unseen prompt forms.

This demonstrates forgetting and a successful rehearsal intervention on a tiny synthetic task. It does not establish broad transfer, durable multi-session learning, or robustness across independent initializations. Next experiments should vary datasets, initializations, update targets, replay budgets and longer sequences of corrections, with matched compute and regression measurements.