SYSTEM1 / THE TEACHING LAB

One small lesson.
A different battle.

Teach when to heal. Keep the move-selection skill. Watch the same encounter run again with fresh local decisions—and compare the result with a single model and a handwritten policy.

Two saved skills · No LLMLocal inference · No teacher calls during playSimulation + optional real Pokémon Red / Blue
CHANGE THE HEALING LESSON

25% HP can be too late.

Keep the original critical-HP rule. Also heal when the next normal hit could knock you out, a potion helps, and attacking cannot finish the battle.

Observation: enough HP to look “healthy”,
but not enough to survive the next hit.

Old label: fight
Corrected label: use_item

Only the healing skill is recompiled.

Inspect the teaching JSON ↗

Measured in the simulator

Recorded evaluation; the battles above run live. Original and corrected policies are both included in the downloadable report.

Survival is a constructed stress test for the HP-threshold weakness. It includes two regressions (8032 and 8055); seed 8004 illustrates an improvement. Broader gains are modest. Review flags and legal overrides are counted; these are not full-game win rates.

Read the results and limitations ↗
REAL CARTRIDGE · CONTROLLED BATTLE STARTS

The same save. One changed skill.

Both players start from identical machine state. Teach healing below, then replay: only the right-hand healing skill changes. Every turn makes fresh local decisions and presses the real game’s buttons.

Checking ROM availability…

Fixtures edit starting HP, inventory, and enemy attack/speed in the starter battle. Those stress stats are artificial; the cartridge handles all later effects. This uses RAM observations, not visual reasoning or campaign play. “starter” keeps the original encounter.

18 fixtures = six conditions × three starting timings. Wins: 6 → 7; strict wins: 2 → 6. Potions: 5 → 20. Two improvements and one regression; threat_slow-37 is an illustrative improvement selected after the full run. Known updated rules match the win count and are faster.

Inspect all real-engine results and limitations ↗