Does a hopping token actually decode? Probe 01 proved the hardware hears ultrasound. This measures whether a detector can read it — per-symbol margin, symbol error rate, and token decode rate against capture time.
Dwell is the experiment. Short dwell = more hops, more null protection, less coherent gain. Long dwell = the reverse. Both phones must be set the same or the listener will never lock.
Loops the token continuously. A 4 kHz control blip sounds once every 8 cycles — audible on purpose, it is what proves the listener's mic path works.
Decision margin is the expected tone against the strongest competing tone in the same slot — what the detector's argmax actually compares. It replaced the guard-band margin as the graded number after field session 1 read 43 dB of guard margin beside 7 % symbol errors: the guard band takes no part in the decision and cannot see reverberant tails or adjacent-symbol leakage, because nothing is ever emitted at 23 kHz. Worst carrier is the highest per-tone error rate — in that session it predicted token decode where both the guard margin and the aggregate error rate did not. Thresholds on both are provisional until recalibrated against decode.
| cycles | capture s | token decode | trials |
|---|---|---|---|
| no lock yet | |||
Majority vote per symbol position across R cycles. This is the number the product lives on: did the whole token decode, and how long did it take? Watch trials — they are overlapping windows, C−R+1 of them, so a short integration makes 0 % and 100 % at R=5 equally meaningless. Raise Integrate s until trials ≥ 5.
| tone | dec med | dec p10 | guard med | errors |
|---|---|---|---|---|
| no lock yet | ||||
Bars are the decision margin per carrier. A tone far below its neighbours is a room null, not a broken phone — and because symbol position → tone is a fixed mapping, a null is a permanently wrong position that majority voting confirms rather than repairs. The guard med column is kept for shape only: its absolute level is inflated, but the reference is identical across tones, so tone-to-tone comparison is still valid. Note that per-tone p10 is the minimum for carriers with n ≈ 7 — one bad look, not a distribution.
Distance is a separate required field on purpose. Probe 01 saved 7 runs labelled unlabelled; Probe 02's first session then saved 5 labelled 1–5, which passed the non-empty check and lost the distance information a second time. Dwell is recorded automatically — do not type it.