Correction: what actually reduced the noise.

The May 23 dev log contains a factual error in its noise reduction claims. This post states what actually happened, and why the distinction matters.

The error

The May 23 dev log makes two claims that contradict each other:

What the post said "The debounce change alone dropped noise from 35% to 8%." — and then, separately — "The filter was implemented and tested against the split dataset. Noise dropped to 8%."

The second claim implies the noise fingerprint filter independently achieved the 8% result. It didn't. The debounce had already brought noise to 8% before the filter was applied. The filter was tested as a proof of concept on that dataset — and showed no significant additional reduction beyond what debounce had already achieved.

The honest picture

Intervention Noise Notes
Baseline ~35% Raw signal, original 5ms debounce
15ms wheel debounce ~8% Single constant change — did the heavy lifting
Noise fingerprint filter ~8% Proof of concept — no additional gain on this dataset

The debounce did the heavy lifting. One constant, changed from 5ms to 15ms on the wheel sensor, dropped noise from 35% to 8%. That is the real result from May 23.

The noise fingerprint filter is a valid design — the concept is sound. But proper validation requires a dedicated calibration ride with no magnet, followed by a learning ride with the magnet, and a comparison of cluster quality before and after. That test has not been done yet. The filter remains in the firmware as a designed component awaiting proper field validation.

The good news

While working through this, Mark II went on the bike. The same session that surfaced this error produced the first real Mark II ride data — and the signal quality is better than anything Mark I produced.

13.2
minutes of ride data, May 30
10,466
PAS pulses captured
18,172
wheel pulses captured
0.254
std deviation on dominant cluster — excellent

The dominant gear cluster locked onto ratio 12.00 with a standard deviation of 0.254. For context: Mark I's best result on the May 23 ride was 0.29. Mark II, on its first real ride, is already tighter. The hardware upgrade is paying off.

Mark II confirmed on cargo bike First real ride completed May 30. Hardware working. Signal quality exceeds Mark I baseline on first attempt. The correction is the setup — this is the payoff.

What comes next

The noise fingerprint filter gets its proper validation test. And the control model moves forward.

  1. Dedicated noise calibration ride — no magnet, then with magnet
  2. Noise fingerprint filter validation against real before/after data
  3. Flat-section clustering (Method A of ensemble detection)
  4. Harmonic pattern recognition
  5. Rider confirmation screen on the Mark II touchscreen
  6. Pre-surgery baseline data collection