Limitations — what has not been shown

The open points, collected in one place instead of hidden in footnotes.

1. SE-M3 detection quality on real images is not evaluated. The detector maps, trains and runs on card 3; the acceptance criteria A1–A15 of the specification have not been demonstrated on photographs of the belt. Everything about M3 on this site is a fit and logic result.
2. One machine, one lab, one lighting set-up. No second installation has reproduced the on-machine figures. All "our observation" rows in Measurements come from a single bench.
3. Small n. Session naming rates are from tens of frames; the 22.7 ms latency is one frame of one scene at an old build and has not been re-measured with the current pipeline.
4. Energy is not measured. The AKD1500 exposes no power meter in MetaTF 2.19.3; energy per correctly recognised object — the comparison the project was started for — is prepared but not published. The historical "22 mJ" figure is not used.
5. Flow-mode timing constants are bench values. 0.25 s stop-for-inspection, 0.6 s stable verdict, 4 s unknown timeout: chosen on the simulated bench, not tuned on the belt.
6. Card order assumption. The reset button maps akida.devices() order to PCIe address order; verified on our carrier, not proven in general.
7. No comparison model. A small YOLO-class detector on the same data and hardware was planned as the comparison baseline and has not been trained.
8. Segmentation figures are synthetic. The 7 % → 98 % improvement and the glare results are from generated scenes designed to reproduce the observed failure; on the belt the effect was confirmed visually, not measured.
9. No formal safety assessment. See guide chapter 12.
10. Alpha software. Interfaces, file formats (vorsa_daten/) and UI can change between builds; the learned-state format is versioned (LERNSTAND) so that old states are relearned rather than misread, but no migration guarantee exists.

Each item links to the place where it would be resolved: 1 → guide chapter 09 and an evaluation set; 2 and 7 → Get SE Inspect (reproductions by others are the plan); 4 → a measurement series with an external power meter on the 5 V input of the carrier.