| Eyes |
The screen model. Given a screenshot and a description, it returns where to click. Used for apps that expose no accessibility tree. |
| Brain |
The decision model. Answers typed multiple-choice questions about one step, with a probability for every option. |
| Hands |
The macOS harness. Reads windows through accessibility, turns each step into questions for Brain, acts, and checks the result. |
| Bench |
Real-desktop tasks with graders that check the final state of files and apps. Suite versions are written v25 and so on. |
| System One |
Our name for the decision style: fast, one-pass choices with probabilities, as opposed to step-by-step reasoning in text. |
/v1/systemone |
The HTTP request format Brain serves: a state, typed questions, and answers with probabilities. |
| choice / score / noul |
The three question types: pick one option; pick a level on a scale; yes or no. |
| Router |
The 0.8B answers first; steps below its threshold, or costly to get wrong, are answered by the 4B. |
| Threshold |
The confidence below which the router hands a step to the 4B. The release uses 0.96, stored as router_threshold in the 0.8B’s deskmind.json. |
| Escalation |
A step answered by the 4B instead of the 0.8B. About 70% of steps escalate in the current release. |
| ASK |
An option the harness adds when the goal matches more than one record; choosing it asks you instead of guessing. |
| DONE / done-check |
DONE is the model saying the task is finished. The done-check makes the harness verify the final state before accepting it. |
| False “done” |
The model said DONE when the task was not finished. The release had none in 39 real-desktop runs. |
| Release tags (G14, G18b, …) |
Training rounds of Brain. G18b is the current release; on Hugging Face its weights are tagged g18b-q8 (8-bit). |
| q8 / 4-bit |
Weight quantization. Brain ships 8-bit; Eyes runs 4-bit in the app. |
| MLX |
Apple’s machine-learning framework for Apple Silicon; DeskMind’s models run on it. |
| Oracle |
A scripted run that knows the correct answer for a sandbox task; used to label training data, never at run time. |
| Held-out / holdout |
Tasks or decisions kept out of training, used only to evaluate. |
| Offline gate |
Before a desktop test, every probe and held-out set is replayed against the previous release; a candidate that is worse on any check does not ship. |
| JevBench |
A third-party benchmark of typed decisions. We report the 231 public items and have requested a sealed run; DeskMind is not affiliated with its maintainers. |
| ScreenSpot-Pro |
A public benchmark for finding UI elements in high-resolution professional apps; used for Eyes. |
