TOD decision
Typed decisions over a state: choices of 2 to 512 labels, yes/no and ordinal scores, with a calibrated probability for every option. Text and up to 8 images per state.
- Model id
- tod-picker
- Endpoint
- POST /v1/systemone · POST /v1/pick
- Price
- 10,000 free decisions per account
- $0.10 input
- No output tokens
Examples
Papers, Please
TOD works the border booth. Each tick it is shown the screen with numbered options, checks the passport against the day's rule, and picks what to do next: approve, deny, inspect or call the next entrant, with a probability for every option. The overlay shows what TOD was shown and what it picked.
Day 3 · 8 min · soundGo to the chair
TOD drives a robot through a scanned house in Habitat. Every step is one /v1/systemone call asking four questions about the camera frame: forward or back, left or right, which way to turn, and whether to stop. It reached the chair in 55 decisions.
Habitat · shown at 3× speed2048
The same one-pass decision read on a small model. Each turn, openjev scores the four moves in a single forward pass of Gemma 3 4B trained with RL on 2048, and plays the most likely valid move. The bars are the model's probabilities, and a toggle switches to the untrained base model for comparison.
Gemma 3 4B · openjev · 2 min