Most filter marketing shows one green column. Excellence hunt v2 ran balanced GSRF Practical against best EMA across multi-family datasets with official Phase5-style metrics. The interesting result is not a single win—it is the split.
Locked scoreboard
| Metric | Wins / N | Mean advantage vs best EMA | Reading |
|---|---|---|---|
| Peak-dev compression | 49 / 52 | ~+37.8% | Green — peaks pulled toward baseline |
| Tracking MAD vs raw | 0 / 52 | ~−412% (much worse) | Red by design — spring costs tracking |
| Oscillation (synthetic actuator mocks) | 30 / 30 | ~+32.1% | Green under predeclare rule |
Family detail (selected)
Oscillation energy (lower better)
| Family | Wins / N | Mean adv |
|---|---|---|
| phase6.2 oscillation mock | 3/3 | ~+35.2% |
| phase5 mock (±noise) | 3/3 | ~+21.5% |
| amplitude sweep A=1…15 | 15/15 | ~+38.0% |
| new regime (chirp / drift+osc / asym) | 9/9 | ~+33.6% |
| VitalDB HR | 6/7 | ~+14.1% (min −6.6%) |
| VitalDB SpO2 | 3/6 | −4.3% mean — not locked as win |
Peak-dev compression
| Family | Wins | Mean peak-dev adv |
|---|---|---|
| All families | 49/52 | ~+37.8% |
| Synthetic excursion | 3/3 | ~+47.9% |
| VitalDB HR | 7/7 | ~+30.6% |
| VitalDB SpO2 | 3/6 | mixed / negative mean |
Why both numbers are true
Peak-dev measures how far filtered peaks sit from a series baseline (median). The spring to \(x^*\) pulls spikes home. Tracking MAD measures how far the filter sits from raw. The same spring refuses to hug raw. If your KPI is MAD-to-raw, you want EMA / Kalman / \(k_{\mathrm{return}}≈0\).
Explicit non-excellences (do not claim the opposite)
- Residual PR-AUC vs labels: 0/7 wins in the hunt framing; residual event detection is not the product
- Spike attenuation ratio: mostly loses vs EMA when “follow the spike” is the metric
- SpO2-flat regimes: mixed — do not sell as universal physiological champion
Verdict
Sell peak calm and mid-band kill. Publish the MAD loss in the same breath. That is how Zero Overshoot earns trust—same posture as publishing server-shutdown risk on comparison sites: show the failure mode that matters.
Sources: excellence_hunt_v2 · EXCELLENCE_FINDING_20260804.md · EXCELLENCE_AGGREGATE.json · /evidence · When GSRF lost