Separate local language-model pilot · Action and notebook generation

This cached SmolLM2-360M policy was trained to generate both an action and the next whole notebook object. Neither adapter yielded a valid command in this evaluation. It has no handwritten game policy, memory compiler, teacher or engine answers during inference. The native seven-room game, legal-action failures, notebook refusal rules, rewards and stopping conditions remain unchanged.

Inspect every registered initial and final run


Positions are assistant turns, including malformed outputs that submit no tool call. Each panel has its own position. A native no-answer stop is separate from an infrastructure censor; an unfinished or unrun case has no scored ability result.

All registered outcomes

WorldAdapterNative actionsNative phaseNative stop / censorScored rewardRaw record

What was trained and what was not established

The base weights were already cached. Both conditions use FP32 compute and the same zero-initial-output LoRA adapter before training. Exactly 168 full native demonstration examples—24 from each room, from the earlier training worlds only—supervise the next move and complete notebook object. Only 102,400 query/value LoRA parameters in the last four layers are trainable. The final checkpoint follows one fixed pass; there is no evaluation-selected checkpoint.

The original two-pass attempt stopped at its registered 06:30 cutoff after 258/336 updates and wrote no final checkpoint. Its unrun post-training abilities remain null. The one-pass condition restarts from the original adapter, with fresh evaluation worlds and a separate registration. It does not reuse that attempt's partial weights.

File/network/process access is denied during decisions. Each prompt contains only the current public screen, static native instructions and declared serialization guidance, and uses a fresh prompt cache. The parser accepts the first complete JSON object and ignores trailing prose. It does not repair an answer or choose a move. Invalid formatting goes through the original empty-tool route. A generated notebook—even valid JSON or a legally accepted write—does not by itself establish correct memory, reasoning, transfer or full-campaign competence.

Independent review and exact artifacts

Initial outcome ledger · Final outcome ledger · Training data receipt · Separate-condition amendment · Original interrupted attempt