WATCH triggered in 27/31 profiles. The average lead time of 96.6s (median 105s) provides ample room for the roaster to intervene. The 4 non-triggering cases share beans with mild RoR fluctuations (Magarrissa-Tulise, Suica types), likely due to the characteristics of the bean type and roasting profile.
Predictions show an average of 25.8s less than the actual remaining time. The cause is the CSD_FC_ESTIMATE_MAX_SEC = 60 setting. While WATCH actually triggers 97s prior on average, assuming a 60s limit translates into displaying an insufficient "60s" right after triggering. Currently, only 16.9% of all ticks fall within "±15s".
Only 12 out of 31 profiles reached the AR(1) ≥ 0.50 threshold. Despite lowering it to 0.50 based on the actual maximum AR(1) = 0.45-0.48 in Jigesa_no17, many profiles still do not reach it. Even when WARN triggers, the lead time is only 20s shorter on average than WATCH, blurring the distinction between the two.
In two profiles, no9 (5s) and no23 (10s), WATCH triggered only 5 to 10s before FC. These also have large positive bias values (overestimation +43 to +50s), indicating the worst-case scenario where FC arrives immediately while displaying "60s remaining". AR(1) may have risen late, or it might be influenced by the warmup period.
Since the WATCH average lead time in actual data is 96.6s, setting the upper limit to 100s will make the estimate at AR(1)=0.35 closer to reality. Sample calculation:
AR(1)=0.35 (WATCH): frac = (0.85-0.35)/(0.85-0.35) = 1.00 → 100s (Current: 60s) AR(1)=0.50 (WARN): frac = (0.85-0.50)/(0.85-0.35) = 0.70 → 70s (Current: 42s) AR(1)=0.70: frac = (0.85-0.70)/(0.85-0.35) = 0.30 → 30s (Current: 18s) AR(1)=0.85 (IMMIN): 0s (No change)
Change location: CSD_FC_ESTIMATE_MAX_SEC = 60 → 100 in live_monitor.py
Checking the maximum AR(1) of all profiles, many plateau around 0.45-0.55. With the current CSD_AR1_WARN = 0.50, more than half do not trigger. Lowering to 0.45 is expected to improve the trigger rate (estimated 60-70%). However, since the differentiation from WATCH will be even thinner, consider reviewing UI representations.
Change location: CSD_AR1_WARN = 0.50 → 0.45 in live_monitor.py
Prediction accuracy is low immediately after WATCH triggers (80-120s before FC, large bias). Errors tend to converge within 30s of FC. Dynamically changing the UI ±15s display to ±40s right after triggering and ±15s close to FC prevents user overconfidence:
AR(1) < 0.50 → "Est. XXs ±40s" AR(1) 0.50－0.70 → "Est. XXs ±25s" AR(1) > 0.70 → "Est. XXs ±15s"
The 4 non-triggering profiles are all Magarrissa-Tulise or Suica types. Beans requiring slow temperature increase may be less likely to accumulate Kalman innovation autocorrelation. Once >10 data points are accumulated, consider adjusting CSD_WATCH thresholds by bean type (e.g., slow beans → 0.30).
| File | FC Time | CSD State | WATCH Lead | WARN Lead | MAE | Bias | Active ticks |
|---|