Separate offline experiment · Actual parameter updates · No model API

A small learner improved early-room control. It did not escape.

A hashed-feature action ranker learned from 2,188 public-interface decisions on 16 training worlds. On eight registered evaluation worlds excluded from training, mean rooms cleared rose from 0 to 1.375. The learner chose every action in the full unchanged native campaign, from reset to its actual stop. No teacher fallback or reference prefix was used during evaluation.

0 → 1.375mean rooms cleared
0/8 → 0/8escapes
9,461of 16,384 parameters changed
12,092nonzero update steps
This is structured-perceptron imitation, not Beam or language-model training. Its recent-board FIFO memory writer is engineered, not learned. The teacher's compiled-plan notebook differs from the learner's memory. A zero-weight initialization is a weak baseline; this small paired pilot does not establish broad generalization, a model ranking, learned notebook use or full-campaign competence.

Inspect the paired full-start runs


Each panel has its own action position because the policies follow different paths. The comparison pairs worlds, not action indices. All failures and native stops remain visible.

All eight evaluation worlds

WorldInitial roomsUpdated roomsInitial rewardUpdated rewardUpdated native stopRaw traces

What was actually checked

Training used 16 separate worlds and 20 fixed epochs. A registration file fixed the source, allocation and settings before world initialization. Actual train/evaluation seed, pair and quartet groups are disjoint. Local registration is not external preregistration proof. Evaluation outcomes were not used for updates, tuning or checkpoint selection.

The trained checkpoint was reproduced exactly from the retained training data. Fresh actor processes and native replay reproduce all 718 initial/updated evaluation commands and transitions. The native game, rules, rewards, 1,024-byte notebook and stop conditions remain unchanged. The learned policy reaches Archive in four worlds, the maze in three, and stops in the first room in one. No learned run reaches Blackout or escape.

Boundaries, artifacts and independent review

Registration · Update record · Reproduction audit · Failure analysis · Policy source · Trained checkpoint · Integration manifest