- Wrong-view customer photos
2026-11-13
14 photos out of training, 11 customers on the partial list, sorter failed its gate, test email with Alex
- Accuracy worklist: tried vs not tried
2026-11-13
Of 58 worklist tasks: 12 done, 6 partly, 40 not started. Plus: was the blind vote fair, and the lane model.
- t50 blind vote: the live model stays
2026-11-13
Alex's 40-pair vote: v39 20, T3-s0 14, 2 ties, 4 both wrong. Live model stays.
- task50 ledger: R1 scorecard, every run
2026-11-13
12 retrains (T0/T3/T1/T1+T3) + v39 reference, all R1 metrics, seeds, costs
- Measuring each stripe on a straightened strip (M1 test)
2026-11-13
New measuring step: precise on clean tape, but sometimes jumps to the wrong edge on real photos. Not ready; Alex's vote decides next step.
- Rule fixes: before and after
2026-11-13
Rule breaks on customer photos 75 to 14 with the fixes on; test score unchanged at 317 of 332.
- Why scans still break the rules we never break
2026-11-13
Audit of 170 live scans: most rule breaks are checker false alarms; the real ones have a fix that is off or aimed wrong. Your 4 photos explained.
- Did our latest retraining beat the live model?
2026-11-13
No new model beat the live one; Alex's vote confirmed: live model stays.
- Lane-detection test (S1), first run
2026-11-13
Lane model finds stripes but places them loosely: 54/327 vs live model 264/327. Parked.
- C12: stripes the model finds but we throw away
2026-11-13
Most missed stripes are drawn by the model but scored under the 0.45 cut-off; 0.15 gives +13 photos on the raw model.
- Worklist v2 checks C9 to C13
2026-11-12
What reviewers see and save (stale cached seeds), plus boom and seed counts. C12 needs one pod run.
- Zero-cost checks C1, C2a, C2b, C3, C6, C7 on the v39 model
2026-11-12
Training recipe, label order, augmentation order, visibility scores, training-set duplicates. Finds flipud reorders points on about 3 in 10 training stripes.
- Model accuracy worklist (v2)
2026-11-12
Reference: Alex's list of checks, builds and retrains to make the stripe model more accurate.
- Accuracy deep dive: where we are and what to do next
2026-11-09
Reference: Alex's write-up on current accuracy, whether more data is needed, and next steps.