Measurements — every figure with its provenance
The master table. If a number appears anywhere on this site, it is here with date, build, hardware, method and limits.
mixed (per row)Event 2026-09-27Published 2026-09-27SE Inspect build 1.9.37
Rows are grouped by kind. "Bench" figures come from synthetic scenes generated by a script in the repository; they show that the algorithm behaves as designed, not how the line performs. "Historical" figures belong to the predecessor system and are kept only in the chronicle.
Measured on the machine (OUR OBSERVATION)
| Figure | Value | Date · build | Hardware / firmware | Method | Limits |
|---|---|---|---|---|---|
| Prototype learning time per example | ~2 ms | Aug 2026 | AKD1500, MetaTF 2.19.3 | edge_learn report field ms_je_beispiel: wall time of fit over all examples divided by count; 48×48 two-channel binary silhouettes into FullyConnected with AkidaUnsupervised |
one card; excludes the Pi-side freeing-up of the silhouette |
| Complete chip learning | 79 examples / 3 objects, step 3 of 3 in 0.3 s | 2026-09-20 · 1.9.36 | 3-card ensemble | learning dialog | includes mapping, excludes photo capture |
| Single-pass core goal, V6 | 1 sequence, 1 pass, 18 / 32 NPs | 2026-08-26 | AKD1500 firmware BC.A1.003.009, MetaTF 2.19.3 | tools/check_single_pass.py, random weights, model.map(device, hw_only=True), mode AllNps; negatives X-gross and X-px320 fail as required |
mapping only — no accuracy statement |
| Hardware limits | input edge ≤ 256 px; input layer ≤ 64 filters; head input ≤ 256 channels; 32 NPs | 2026-08-26 | as above | error messages of model.map |
bound to this firmware |
| Camera → verdict | 22.7 ms | 2026-08-28 · 1.2.6 | Pi 5 + 1 card | acceptance of the touching-parts scene (SD card + carabiner); the UI figure "Inferenz" at that build measured the whole path incl. Pi-side segmentation | one frame, one scene; today's pipeline not re-measured |
| Session naming rate | 100 % (26/26); later 98.5 % | 2026-08-28 · 1.2.6 / 1.3.9 | as above | recognition log of one session | small n, laboratory scene |
| Card hang and recovery | DMA wait completion timed out every ~5 s on 0001:06:00.0; recovered by remove/rescan in < 10 s |
2026-09-20 | 4 cards behind PCIe switch | dmesg, manual sysfs remove/rescan |
single incident |
Bench (synthetic; scripts in tools/)
| Figure | Value | Date · build | Script | Scene | Limits |
|---|---|---|---|---|---|
| Reference pipeline without a network | centre error median 0.19 mm, angle 0.54°, ~15 ms per frame | Sep 2026 | tools/run_selftest.py --use-background |
five synthetic scenes (free, touching, overlap, mixed, stacked), perfect calibration | not on the belt |
| Belt model vs. vignetting | card IoU 0.98 (old method 0.07); empty belt 0.0–0.7 % false foreground; part at the border 0.95; 45 % part 1.00 | 2026-09-19 · 1.9.33 | tools/test_farbmaske.py |
dark-green belt with parabolic, cos⁴ and flat vignetting, noise σ = 3, black card with white label | synthetic |
| Glare filter | empty belt with three flickering spots: 0 parts (with and without reference); card next to spots 0.97; card over a spot 0.95; large white part 0.98 | 2026-09-19 · 1.9.35 | tools/test_glanz.py |
saturated cores with soft halo, 25 % flicker | synthetic |
| M3 head logic | 8 hits with slot-aware Soft-NMS vs. 4 with standard NMS; assignment Slot→Objekt [1, 0, 2] |
Aug 2026 | tools/test_m3_head.py |
three locally overlapping objects | logic only |
| Line state machine | flow and single modes, unknown handling, arm cycle, total-failure path | 2026-09-19 · 1.9.33 | tools/test_linie_fluss.py |
simulated detections, simulated time | timing constants not tuned on the belt |
| Card watchdog | failure, lock-out, retry, rejoin, total failure | 2026-09-20 · 1.9.36 | tools/test_karten_waechter.py |
fake cards | — |
| Augmentation angle test | 30 cases pass | Aug 2026 | vorsa.augment.pruefe_winkelmitfuehrung() |
— | found a sign error at 45° during development |
Historical (ENGRAM, LinkedIn series, August 2026)
| Figure | Value | Note |
|---|---|---|
| Endurance run | 8,238 detections in 10 minutes, zero dropouts | after PCIe Gen 1 / ASPM-off fix; not reproducible from the current repository |
| Early prototype learning | 75 % hit rate over 5 classes with a random-weight backbone, num_weights 48 |
predecessor method |
| "22 mJ per recognition" | shown in a Day-5 graphic | method unknown today; not a current claim |
Explicitly not measured
AKD1500 power: MetaTF 2.19.3 exposes the power meter only on the AKD1000; on the AKD1500 device.soc is None. Community figures (< 1 W per card) are third-party and are not used as our numbers. Energy per correctly recognised object: prepared as a measurement series, not published.