System1 / experimental skill composition

Teach a piece. Change the behavior.

Three small local skills combine to collect a key, open a gate, and bring supplies home.

● Live local inference · No LLM
The world changed. The old lesson did not.

Purple used to be safe. Run the mission, then teach only the terrain skill and try again.

Inside the System1 skill network

1 · ObjectiveInventory + gate + positionAwaiting observation
→
2 · TerrainAdjacent surface observationsOriginal lesson
→
3 · MovementChoose among eligible movesAwaiting decision
Each step calls three saved System1 skills. The result changes the world; the next observation starts the loop again.
CourierKey / gateSuppliesLavaPurple tile

The terrain skill excludes moves it marks dangerous or uncertain. The wiring selects the highest remaining score, or stops if none remain. No path planner or world-truth shield runs here. Other review flags are shown; this simulator still executes those predictions.

A correction you can read

Teach a surface decision. The objective and movement files stay unchanged.

{"input": "surface_purple light_day texture_rough", "label": "dangerous", "split": "teach"}

8 new teaching labels + 22 separate calibration labels. Obsolete purple lessons are replaced; other terrain lessons are retained. Calibration labels cost teaching effort too.

Recorded evaluation, separate from this live mission
Loading measured results…

Full evidence JSON · Read the teaching examples