← back to test matrix

Test 1.1

Can we recover missed stripes for free by running the model multiple ways?

Net-negative as tested. We recovered 35 missed stripes, but that cost 132 new false lines (66 duplicates + 91 nowhere near a real stripe) — more junk than it's worth as-is.
What it means for SailScan

This was the cheap test: if the model truly never learned faint stripes exist, no amount of re-running it differently should conjure them — and that's roughly what happened. The weak recovery, even with 5 extra passes, is itself evidence the miss is a LABEL problem (the model was never taught these stripes), not something more computation alone can fix. Points toward the label-audit work (test 0.2) as the real fix.

+35
stripes recovered
+132
new false lines
net-negative
verdict
The detail
PassRecallPrecision
Baseline (rules stage)94.8%98.1%
+ TTA, 5 passes merged (rules stage)98.0%90.0%
LedgerCount
Stripes recovered that we were missing35
New false lines added132 (66 duplicates of a real stripe + 91 nowhere near a real stripe)
Real stripes lost0

We ran our existing model 5 different ways on each photo (different zoom, a zoomed-in bottom crop, a lower confidence bar) and merged what came back. No retraining, no new model.

Source: 332-photo holdout, frozen 2026-08-25. 5 TTA passes (base/960/1280/bottom-crop/low-conf), custom polyline fusion.

raw data (JSON) →