← back to test matrix

Test 0.8

Do our stripes ever double back on themselves — could CLRerNet even work?

GREENLIGHT. Only 1.28% of stripes double back by more than jitter — comfortably under the 2-3% go/no-go bar. CLRerNet's core assumption holds for our data.
What it means for SailScan

CLRerNet is a lane-detection model built for painted lines on roads — a close match to sail stripes. But it can only draw a stripe as one x-value per row, so it would break on a stripe that curls backward. This test proves that almost never happens on real sails. Test 3.2 (the actual CLRerNet fine-tune) is cleared to proceed.

1.28%
stripes that double back
GREENLIGHT
verdict
332
photos audited
The detail
TierMeaningViolation rate
Tolerant (primary, jitter-safe)Backward step > 1% of the stripe's own span1.28%

CLRerNet (a self-driving lane-detection model, test 3.2) can only be used if every stripe is "single-valued" — never doubling back on itself once you look at it sideways (luff→leech). This test checked all 332 holdout photos' labels for stripes that curl back. Only 1.28% do, well under the go/no-go bar of 2-3%.

Source: 332-photo holdout, frozen 2026-08-25. Geometry-only check on stored label points, no model run.

raw data (JSON) →