Experiment 004 / Compete the evidence

StateMatrix

Five gesture families enter. Time, margin, and ownership decide whether one leaves as intent.

Camera starts only after click. Frames, landmarks, and decisions stay inside this browser.

Continuous evidence

Five scores

  1. 01Pinch 0.000
  2. 02Fist 0.000
  3. 03Open palm 0.000
  4. 04Point 0.000
  5. 05Two-hand span 0.000

Discrete decision

Why this state

Phase
UNKNOWN
Winner
-
Margin
0.000
Owner
-
Timer
-
Reason
evidence missing

No evidence selected

Recent transitions

Recent transitions

    Ready. Camera permission has not been requested.

    Continue research spine Experiment record / 004

    Experiment record / 004

    Competition before certainty.

    Phase
    Compose vocabulary
    Study status
    Physical validation pending
    Decision
    Revise
    01

    Question

    Can five gesture families share one temporal contract?

    02

    Hypothesis

    Normalized pose scores plus a winner margin should reject overlap before temporal confirmation turns evidence into intent.

    03

    Deterministic evidence

    • Pinch, fist, open palm, point, and span confirm after 140 ms.
    • Short holds never activate.
    • Direct handoffs pass through release and cooldown.
    • Competitive evidence remains unknown.
    • Span requires two open hands.
    04

    Timing contract

    Confirm
    140 ms
    Release
    100 ms
    Cooldown
    180 ms
    05

    Not yet proven

    Real hands under blur, occlusion, changing light, crossings, sleeves, jewelry, and varied devices still require physical measurement.

    06

    Decision / revise

    Keep explicit competition. Test ownership with real crossings.

    Deterministic evidence supports one shared state machine. Threshold reliability and identity behavior remain provisional.

    Method / 004

    Evidence earns intent.

    The detector does not name a pose from one frame. It measures five competing, palm-normalized signals, rejects overlap, then confirms one owner over elapsed time.

    01

    Signal path

    Landmarksup to two hands Five scorespose + span ratios State gatehold, release, cooldown

    Palm scale makes pinch distance and two-hand span comparable across hand sizes. The strongest score still loses when runner-up evidence is too close.

    02

    Competition before certainty

    Pinch0.28
    Fist0.14
    Open palm0.61
    Two-hand span0.83
    const [winner, runnerUp] = rank(scores);
    const margin = scores[winner] - scores[runnerUp];
    const eligible = scores[winner] >= 0.78 && margin >= 0.16;

    Activation needs score ≥ 0.78 and margin ≥ 0.16 for 140 ms. Continuation is easier, but direct handoff still requires release and cooldown.

    03

    What gets rejected

    • Missing landmarks, impossible geometry, or palm scale below 0.02.
    • Competitive evidence: strong winner, insufficient margin → unknown.
    • Identity change, evidence gap, or short hold restart temporal confirmation.
    04

    Implementation map

    src/gesture/gesture-state-matrix.ts

    Measures normalized scores, ranks evidence, runs temporal state machine.

    src/engine/state-matrix-fixtures.ts

    Produces deterministic ambiguity, handoff, crossing, and span evidence.

    src/effects/gesture-matrix-effect.ts

    Consumes confirmed signal only; never tracker internals.