Start with quality.
Freshness, source availability, and timing integrity describe how much evidence was actually observed.
OBSERVATION · TYPING · MOUSE MOVEMENT · CLICKS
Measure the interaction.
Keep the evidence clear.
Explore keyboard timing, pointer movement, clicks, and scrolling in one local task. Start a capture and watch recent windows update while you interact. Turn Sliding window off to measure one complete capture.
Ready to capture0.0 s
Checking classifier availability…
Start to collect a 10-second window, then update every 2 seconds.
No window has been assessed.
Input collection continues while a window is measured. Slow updates use the newest available window; they do not queue old windows. Automatic stop is scheduled after 120 seconds. Classification availability is shown above.
Record Additional Behavior Measurements in each window. These measurements are not trained into a classifier yet. FP-Agent is reserved for evaluation; the previous classifiers are disabled while replacement training is pending.
The optional probe observes selected Element method calls inside the task. It excludes arguments and DOM contents and cannot see every inspection path.
OPTIONAL CAMERA MEASUREMENT
Approximate gaze location
Centre your whole face in the preview. Adjust the camera until your forehead, eyes and chin are visible, then keep that position. The camera will stop after 30 seconds if you do not continue.
Enable the camera, check your position in the preview, then select Start calibration dots. Look at nine calibration dots, followed by five check dots. Camera processing stays on this computer. When debug recording is enabled below, calibration video and diagnostics are saved locally. Gaze shows where you may be looking; it does not prove who performed an action.
Gaze model: WebEyeTrack BlazeGaze, a pretrained neural network using eye images and head pose. Your nine calibration dots personalize its screen mapping; five separate dots check accuracy.
You can select this now. The point appears after calibration passes, without starting a capture. Gentle display smoothing reduces jitter with a small delay; accuracy checks and exported measurements use the unsmoothed estimates. Selecting this does not turn on the camera.
Enabled for debugging. Records the camera view during calibration and any retry, with no audio or screen capture. Video, dot timings, predictions and errors save automatically to the HCA folder on this computer; nothing is sent to an external service. Recording ends when calibration passes, stops or fails. Keep the tab open until Saved appears.
Debug recording is ready for the next calibration. The camera is off.
Camera off. Gaze is optional.
Not calibrated. Calibration is approximate and depends on lighting, camera position and how steadily you look at each dot.
No current gaze estimate.
If the first accuracy check fails, you can retry its five dots once. The camera waits for up to 30 seconds, with no gaze inference during that wait. Finish, Clear, leaving this tab, resizing the window or Stop camera releases it. Camera processing adds load and may affect responsiveness measurements.
Only explicitly started activity inside this task is captured and processed locally. Typed text, key identities and DOM contents are excluded. Measurement files save when Download is selected; enabled camera debug recordings save automatically on this computer. Automatic stop is scheduled after 120 seconds.
Choose who actually performs the next capture. The label is saved in downloads, never used to decide the result. It is fixed when Start capture is pressed. Use an opaque tester code rather than a name, so later evaluation can keep each tester’s sessions together.
TRY THE TASK
A few natural interactions
Move between the targets, select them, and scroll through the panel.
No target selected
Scroll to explore
Overlap is a measurement.
A key can be pressed before the previous key is released. Signed intervals preserve that overlap.
Time matters.
Movement speed uses actual elapsed time. A stationary path has zero distance and undefined efficiency.
Scroll back, then forward.
Scroll direction changes and target geometry add context to the task.
LATEST WINDOW
Your measurements
Evaluation data includes this test’s timing records, pointer coordinates, and summary for local testing. Typed text and key identities are excluded. A measurement file is saved only when you choose Download; no human or agent label is assumed. Enabled camera debug recordings save separately.
Complete a capture to see measurements and the optional local assessment.
Evidence captured
Detection sample requirements are not yet validated. Calculable measurements alone do not establish enough evidence to identify input origin.
EXPERIMENTAL LOCAL ASSESSMENT
Model estimates are not validated human-control confidence. Action authorization is separate.
Automation includes scripts and AI agents. Controller type is not observed by this browser test.
LOCAL CLASSIFIER RESPONSE
Raw response
No response yet.
No response yet.
Recent windows
Each row describes its recorded interval. Overlapping windows share evidence and are not independent tests. The latest 60 outcomes are retained.
| Window | Result | Human / automation | Input | Latency |
|---|
Additional Behavior Measurements
No additional measurements yet.
Unavailable values reflect missing, unsupported or boundary-cut evidence. These measurements do not identify the operator and are not trained into a classifier yet.
Group and measurement names explain each value. Hover over a measurement name to see its original data key; the full keys also remain in the downloaded summary.
Inspect the 74 additional measurements
Gaze measurements for this window
No gaze summary for this result yet.
Measurement downloads contain numeric gaze summaries. Debug recordings are separate local video and diagnostic files when enabled above. Facial landmarks and calibration weights are never saved. Missing gaze is not evidence of automation.
Gaze–action coordination
No gaze–action measurements for this result yet.
These measurements connect gaze available before each action to its selected target. The tolerance around a target comes from calibration error; nearby targets may be ambiguous. A person can watch an agent act, so target alignment does not identify the executor. These features are collected for future classifier training and do not change the current classification.
Inspect recent actions (up to 20 shown)
Inspect the feature summary
No snapshot yet.