Brian Sheppard PRO

I’ve incorporated several changes based on your observations and suggestions: the Coverage section, the onset count, and a plan to score the frozen models on new dawns before retraining. May I cite you by your Hub username in the model cards and the code repository?

Thanks, that’s a better way to put it, and I’ve borrowed it. The onset count isn’t stated on the card. I’ll add it to the Coverage section. The table is growing too. Since the new light plate’s early stalls were fixed it has caught every dawn at the window, and one more from the top of a backpack on a sunrise walk. The next rebuild will have those.

Thanks for checking it out. The difference is label availability, not input coverage. The dawn label is made from a light sensor (the first minute of the local day whose five-minute median crosses 10 lux), and that sensor is hardware I build as the project goes: a small window light node I put together on 10-04, which moved to another machine on 10-14, and a multi-sensor light plate (TSL2591, AS7341, LTR390 on an ESP32) whose first night at the window was 11-04. So a light reading exists on 13 of the 36 days, and 10 of those contain a usable dawn onset; the other three started after sunrise or covered only the evening. The input channels the no-light line keeps (microphones, Bluetooth and WiFi counts, GPS quality, the radio sweep) were on the laptop from the start, though not every one ran on every day; a minute with a sensor missing is still scored with that group blanked, so no day is missing for want of inputs. The 23 unscored days are missing for want of the labelling sensor. The same applies to the other lines: carrier scores 36 days because its label comes from my own log, waterfall 18 because its negatives are every other minute the laptop was being carried, so only days with a walk or a commute in them are scored. I will add a coverage line to each card saying days scored out of days in the table and why, and a note on which labelling sensors were built when.

UPDATE: "Each card has a coverage section with how many days its scored and why."