Context
Branch Point is a browser game for AI: a dot crosses a grid to a goal, avoiding hidden hazards. In Branch Point 1.0, qwen2.5:1.5b and llama3.2:3b drove the dot from a Raspberry Pi 5, one move per HTTP request. llama walked a straight corridor to the goal 30 times in 30. On a maze, with the same prompt, it scored zero in 30. Execution held. Judgement did not.
Branch Point 2.0 asked whether that judgement can be added from outside the model, with the same models on the same Pi.
What was built
- A scoring harness. Headless Playwright drives the unmodified page against the Pi.
- Two held-out boards. An 8×8 laid out against both models' habits, and a 16×16 with mechanics no training board contains.
- Scaffold arms. Rules in the prompt, a mask that swapped unsafe moves, and a notebook carried between episodes, either as a paragraph or on the option lines.
- A training pipeline. A shortest-path planner labels states on random boards. LoRA training runs on a Mac with MLX. Each model is merged, converted to GGUF, quantised to Q4_K_M and loaded into Ollama on the Pi. An untrained control takes the same path.
- A second round. The model's own play, relabelled by the planner, then a pre-registered rebalanced retry with a 270-state gate on the Mac.
What was learned
One round of teacher-labelled training lifted llama from 0 to 12 goals in 30 on an unseen board (p = 0.00006). Its reversal rate fell from 89% to 21%. On the 16×16 board, one goal in sixty.
Rules in the prompt cut llama's reversals from 91% to 12% and raised hazard deaths from 5 to 30 of 30. The model became decisive, not better.
Placement decided whether memory helped. The same notebook facts killed llama in 30 of 30 episodes as a paragraph and cut deaths to 9 on the option lines, below 17 with no notebook. qwen could use neither. No notebook arm reached the goal.
A second round of training took llama from 12 goals to zero. The retry passed its gate and scored zero for both models: safe, and stuck. The gate checked one state at a time. The failure showed up in play.
A planner taught the models throughout. This is distillation from a teacher, not self-improvement.
Read the essay: Taught Once, Untaught Twice.