Back in early September we tried using SAM3 to redraw Deep20's stripe lines by pointing at them. It made the lines worse on most stripes, a closed negative result. This test asks why: did SAM3 get too weak a hint (a few points), or is it just bad at this job regardless of the hint? A fuller prompt (points, a rough box around the stripe, a rough mask) could turn a weak hint into a usable one. If SAM3 still fails with all that help, that kills the idea for good.
We checked, without touching a GPU, whether SAM3 even accepts a box and a mask alongside points. It does. We traced the actual code path line by line rather than trusting the claim on its face. The richer-prompt idea is real and testable. What's missing is the measurement itself: run both versions on real photos and see which one draws closer to a human's line. That run needs a rented GPU. None was on, so the numbers aren't in yet.
| Question | Answer |
|---|---|
| Does SAM3 accept points + box + mask together? | Yes. Traced the real code path used when our sam3.pt weights load: it's a point/box/mask-promptable architecture (not the separate text-only variant), and a supplied mask genuinely reaches the model's mask decoder rather than being silently ignored. |
| Was inference actually run to compare points-only vs. the richer prompt? | No. We checked every prior GPU pod for this project live; all were off. Starting a new one wasn't authorized for this investigation. Running SAM3 locally caused a prior crash, so we stopped instead of risking that again. |
| Is the test harness ready to run? | Yes. The predict script, pod launcher, and scorer all exist and are reviewed. Once a pod is up, this runs end to end on the same 332-photo frozen holdout used for every other model eval. |
| What was the original points-only result being re-tested? | SAM3 redrawing every Deep20 stripe with points alone landed at 36.4px average error vs. a human line, catching only 40% of stripes within tolerance. Deep20's own lines score 23px error and catch 91%. |
"Points-only" and "points + box + mask" in this test mean: Deep20 finds the stripes first (it already follows our sail rules), then SAM3 is asked to trace each one more precisely, either from a few clicked points alone or from those points plus a loose box around the stripe plus a rough guess-mask. We check whether the redrawn line lands closer to what a human traced.
Frozen 332-photo holdout (test-holdout-ids-canonical.json, frozen 2026-08-25), never used for training. Points-only baseline number is read from the 2026-09-02 run's own recorded results, same scoring method the richer-prompt test will use. Test-only investigation; nothing shipped, no production code touched.