Your recordings stay on this computer
Load saved captures, train your own model, and review its suggestions. No account, upload, pretrained weights or external service is used. Original files and labels are never overwritten. This is an experimental foundation; later original sensing firmware has a separate, unverified protocol.
No recognition accuracy is established without real labeled recordings. Activity states and transition events are separate models. Three-second causal windows add at least three seconds of context latency. Human-click event timing is approximate. Workflow and limits
Load a private capture or feature dataset to begin.
1 · Prepare recordings
No data loaded.
2 · Assign independent captures
Choose whole takes, never adjacent frames. Training requires all classes from at least two independent captures per class. Reserve different takes for held-out evaluation. Changing names does not make a copied recording independent.
3 · Train, save or load a model
No trained model. Model files contain local learned parameters, not executable code. Saved offline models sort loaded recordings; they do not replace the live tab’s separate preview model.
No held-out evaluation yet. Unknown/rejected cases count toward coverage limits.
4 · Human review, separate from original labels
Proposals are suggestions, never ground truth. View the original capture/context before assigning a label. Choose event “none” only after explicitly reviewing that whole window. Quality-rejected windows cannot be forced into training with a label. Keep held-out annotation blind to predictions when possible.
All exports are separate derived files with hashes tying them to the inputs. Keep original captures. There is no automatic denoising, camera synchronization, event boundary detector, medical use or safety guarantee.