Repeated exposure for learning in an added shard

The short capacity experiment gives each fictional fact twelve training targets: eight questions and four narrative descriptions. Its first completed candidates learn the JSON format but recall few of the facts. This experiment isolates training duration as a possible cause. It changes neither the target answers nor the required learning and retention margins.

The plan repeats the original 128 batches eight times at the original lower peak learning rate, for 1,024 updates and 96 target exposures per fact. The cosine schedule spans those 1,024 updates. Both appended and existing-tail arms keep the same optimizer, microbatches, padding, reference KL, margin penalty and proven-trained replay examples. Only the terminal checkpoint can qualify.

Knowledge Capacity Scaling Laws reports that repeated exposure substantially affects fact learning in its controlled pretraining experiments. That motivates measuring duration here; its exposure counts are not a proven threshold for our frozen-parent setting. Knowledge Storage and Extraction also distinguishes memorizing text from answering different questions about it. The existing generated-answer gate therefore remains necessary.

Training and memory ownership

The distributed prefix produces a content-bound bank for the exact 128 training batches. Every stored microbatch binds its parent, source/profile, input order, labels, loss weights, padding and tensor hashes. Identical prefix and reference representations share storage. Subsequent updates load only the current batch and train the owned tail with a read-only output head.

Established owners keep their assigned layers. Their weights and previous Adam history remain unchanged. The new owner holds its tail, its own Adam state, one batch of representations and the frozen output-head replica. No owner needs the whole model. This is an operated experiment; permissionless feature certification and payment are separate work.

The midpoint at update 512 preserves the complete tail and optimizer cursor for recovery. A resumed worker recreates the immutable bank from the same parent and inputs, then continues the bound schedule without resetting Adam. All owners commit the terminal checkpoint before any development scoring.

Freeze and selection

Preparation refuses to proceed until two prerequisite records are committed:

• Exact real-model feature factorization for both arms, including all eight intermediate updates under the declared source commit.
• Failure of all four original short-run candidates, with neither arm selected to open the original final set.

The prepared record binds those receipts, every numerical dependency, the existing tokenized inputs and the repeated batch cursor. Training refuses uncommitted or changed preparation. If the original experiment succeeds or feature factorization fails, this plan cannot run as written.

Both rehearsal arms must finish before selection. Eligible terminal candidates must pass every existing development gate; finals use the still-unused knowledge questions and the previously exposed retention probes. A failed arm cannot open finals. Learning in added blocks and an advantage over the fixed-size control are reported separately. No numerical result in this experiment issues tokens or changes serving.

Validation

python -m pytest -q tests/evolution/test_feature_bank.py tests/evolution/test_rehearsal.py

The three-process integration checks exercise feature production, durable storage, repeated local updates, checkpoint writing and answer generation for both arm layouts. Their final parameters and Adam states match eight ordinary distributed updates exactly. Additional checks reject altered feature bytes, rehashed replacement banks, changed input order, incomplete controls and premature checkpoint selection. Failed development does not read final inputs.

Real-model feature equivalence passed all eight updates in both layouts. The prepared record and bounded execution plan bind the longer trial before any worker allocation. Its learning outcome remains pending.