Experiment 003 / Compare the proof

CalibrationBench

One hand. Three pipelines. See whether filtering and a private session profile make pinch and fist evidence more truthful.

Camera permission appears only after click. Calibration keeps scalar geometry in memory and discards it on reset.

Gesture Evidence

Skeleton Mirror

Calibration sequence

Not started Open → pinch → fist. 700 ms evidence each.

Same observation stream

Three claims

Fixed

UNKNOWN

Metric: -
Gate: -
Errors: -
Latency: -

Filtered

UNKNOWN

Metric: -
Gate: -
Errors: -
Latency: -

Calibrated

UNKNOWN

Metric: -
Gate: -
Errors: -
Latency: -

Ready. Camera permission has not been requested.

Continue research spine Experiment record / 003

Experiment record / 003

Same hand. Different truth.

Phase: Name intent
Study status: Active validation
Decision: Revise

Question

Can local calibration improve two gesture families?

Hypothesis

A private session profile should reduce threshold errors for pinch and fist without adding more than 80 ms of deterministic transition latency.

Comparison

Fixed: Shared gates
Filtered: One Euro
Calibrated: Session medians

Observed failure

Fixed gates miss deliberate personal-range evidence.
Filtering reduces jitter but cannot repair a wrong gate.
Overlapping calibration references return inconclusive.
All calibration geometry disappears on reset.

Measurement contract

Latency budget: ≤ 80 ms
Filter beta: 1.2
Capture: 15 samples / 700 ms

Deterministic evidence only; the physical test matrix remains open.

Decision / revise

Keep the three-way comparison. Move the claim into real hands.

The bench makes pipeline differences inspectable. Calibration remains provisional until physical camera tests establish error-rate and latency benefit.