No request sent yet.
Probabilities are conditional option scores: use them to rank options and set coarse thresholds, never as calibrated confidence.
scoring…
GPU work serializes on a per-request lock — concurrent requests queue, they do not speed each other up.
Level
Manual drive (when idle): →/D run right, space/W jump.
Physics pauses while a decision is pending. In Guided mode the state text includes a hint about the right move;
in Unguided mode only the facts are given and the model has to decide on its own;
in Run-length mode the terrain is a compact run-length encoding
(-N floor, #N pit) the model must parse;
in ASCII mode the facts are a single symbolic row (- floor, # pit, * you, ! flag)
the model must decode — the hardest of the four.
⚄ Randomize moves the three pits to new (always-solvable) positions; the seed box shows the seed so a
layout can be reproduced — type a seed and press Enter.
Level
Same physics as the Platformer demo, but the decision model is given no rules — only a
neutral observation of the level and the actions run / jump, plus an
insufficient escape hatch. When it picks insufficient with p ≥ 0.99, or the player
falls into a pit, or it keeps requesting jumps with none left, the client calls /plan: a reasoning
chat completion that reviews the transcript of completed actions and writes a set of rules. The rules then lead
every later observation — that is the only way the model learns how to play. Every trigger restarts the level
with the learned rules: the planner writes full game rules, not a fix for one stuck spot, so the model gets a
fresh attempt to apply them from the start. Reset restarts the level and keeps the learned rules;
✕ Forget rules clears them. Manual drive (when idle): →/D run right,
space/W jump.
Room
Manual drive (when idle): ←/A turn left, →/D turn right, ↑/W step forward.
The first-person view is yours — the model only ever sees the text in the side pane: the four adjacent cells and what is in view,
so it must build its own map by dead-reckoning. It must find the key, unlock the door in the dividing wall, and reach the exit,
one action per decision, within the step budget. In Guided FPP the state includes a compass hint toward the key/door/exit;
in FPP only what you see and what happened; in Map it gets the whole layout.
⚄ Randomize builds a new (always-solvable) room; the seed box reproduces it — type a seed and press Enter.
Room
Same room as the Puzzle room demo, but the decision model is given no rules — only a neutral
observation (all four adjacent cells, in-view bearings, carry status and a step count — or the whole top-down map)
and the actions forward / left / right, plus an insufficient
escape hatch. When insufficient is its leading option at all, or the step budget runs out, the client
calls /plan: a reasoning chat completion that reviews the transcript of completed actions and writes a
set of rules. The rules then lead every later observation — that is the only way the model learns how to play.
Either trigger restarts the room with the learned rules: the planner writes full game rules, not a fix for one
stuck spot, so the model gets a fresh attempt to apply them from the start.
Reset restarts the room and keeps the learned rules; ✕ Forget rules clears them. Manual drive (when idle):
←/A turn left, →/D turn right, ↑/W step forward.