The questions worth asking.
Illustration of calibration geometry: fixation targets at known positions. Not data.
Five questions the channels have to answer.
- 01
Contact force
What is the held-out-user error after a 15-minute calibration?
Stand-in forearm sEMG/IMU bands, calibrated against fingertip force sensors, are to estimate per-finger force. Ember, the wrist band, is to replace them after its build gate. Cr4dle holds no force data that it recorded from a person. The project on the public Hyser corpus closed on its kill rule. On Hyser, the band emulation did not beat its zero-calibration pipeline or a constant predictor after calibration. On putEMG-Force, the 24-channel pipeline beat the constant predictor but not its zero-calibration pipeline. The number that counts is the cross-user (LOSO) error against a learned vision baseline, printed on the card: pooled 1.5 N or better at the 2026-10-05 gate, whose date passed with no result recorded in the registry, or the channel thesis is in question.
- 02
Gaze target
Does natural gaze earn its place in a policy?
Published gaze-conditioned imitation results are replicated on a research task under a held-out-user split, and the number is published either way. A gaze target is where the wearer looked: a point on the scene camera, on the video clock; it is not a mental state.
- 03
Sync error
How far apart are the clocks, really?
An LED flash the camera sees and the band logs gives each stream an offset; the residual is the sync error, floored at the detectors' resolution and re-derived from the raw area. It is measured on every recording, never copied from a spec sheet.
- 04
Demonstrator state
Does a physical state predict a discarded episode?
Fatigue and safety-relevant physical states are a column on every episode. The open question, with a kill date, is whether that column predicts discards earlier than jerk and completion time. It never claims emotion, mood or stress.
- 05
Social navigation
Does the wearer's gaze add to head yaw in predicting the wearer's next path, in staged walks past another person?
Lane two is a separate research lane: consented people are to walk among other consented people in staged sessions, recorded from the wearer's side. A frozen pre-registration checks the pipeline in a simulator first, with null, echo and planted controls, and that check never stops the lane by itself. It tests the pipeline, never gaze in people. A failed sensitivity part of that check makes a stop reading inconclusive, and a failed specificity part makes each recorded reading inconclusive. No lane-two encounter is recorded, so no result exists. Gaze is an input only, never a model target, and no model forecasts a bystander.
- Current work
Make the kit inspectable.
The recording path, the redaction stage, the exporter and the consent gate run end to end on synthetic data (`make demo`) and print the shape of every artefact, including a card whose unmeasured numbers say not run with their gate.
- Date passed · no result recorded
Put the kit on a person.
The date of the 2026-10-05 gate (sync error printed, at least ten hours from five participants with a card, and a cross-user force model against a vision baseline) passed with no result recorded in the registry. A new date is a founder decision. The next gate is the force ablation on by-hand episodes with at least 40 trials per arm. ADR-0136 suspends the paid gates of this phase. The research gates stay in force.
- Our standard
Let the held-out number set the scope.
A channel ships under a label-validity spec that names what grounds it, on whom it was validated and its evidence kind. A synthetic or simulated number is labelled as such and is never release-eligible.
A label is
not a story about you.
We do not claim to read thoughts, memories, dreams or emotions. Demonstrator state names fatigue and safety-relevant physical states under a label-validity spec, never a mental state; physiology such as EEG, EDA or heart rate is a recorded research column, never a product.
The EEG research tiers record research-only columns in bounded sessions on bought equipment, each question registered with a kill date before any session. Progress on one channel does not establish another; each earns its place with its own held-out-user number.