Silent control for complex tools
Decode low-bandwidth intent into cursor, language, and robotic actions without forcing users into a fixed command vocabulary.
Research / 05
Evidence before
autonomy.
We study the point where a neural model should act, wait, ask, or admit that it does not know.
Related case →
Models that know when not to act
Estimate uncertainty, fatigue, and task context so the system can abstain, ask, or wait instead of confidently guessing.
Interfaces that learn both ways
Study how people and models co-adapt over time while preserving agency, reversibility, and an observable chain of decisions.
A good neural interface is not the one that predicts everything. It is the one that makes uncertainty visible.
← Back to the main systemResearch map / 06
Four questions guide the lab.
What survives a new day?
We measure how personal adapters preserve intent when sensors move, fatigue changes, and the task becomes unfamiliar.
Can the model show doubt?
Confidence is treated as a first-class output so people can see when the system is ready, waiting, or wrong.
Who owns the final action?
We design reversible loops that keep the human in control of the policy, the context, and the moment of commitment.
How does a signal become useful?
We evaluate real workflows, not only benchmark accuracy: cursor control, communication, robotics, and adaptive tools.
Evidence stack / 07
What we measure before we call it intelligence.
01 / SIGNAL QUALITY
Can we trust the input?
Artifact rejection, channel dropout, contact quality, and timing drift are logged before any model score is reported.
02 / DECODING
Does intent survive?
We compare within-session accuracy with cross-session transfer, task changes, and deliberate uncertainty.
03 / HUMAN FACTORS
Does the person stay in charge?
We track workload, fatigue, correction rate, and recovery time—not just whether an action was technically correct.
04 / DEPLOYMENT
Can it live outside the lab?
We test calibration time, sensor movement, connectivity loss, and how gracefully the system abstains.
Protocol stageoffline replay / guided task / open-world taskgated
Primary metricaccuracy / calibration / recoverymulti-axis
Failure reviewhuman correction + model tracerequired
Release ruleperformance plus agencyearned
The first ten minutes matter.
A system that needs perfect setup is not adaptive. Our calibration work studies how quickly a model can become useful while showing exactly what it still does not know.
Silence is a valid output.
When signal quality drops or intent is ambiguous, the safest action may be to do nothing. We treat abstention as a capability to be designed and measured.