Verification Bench

Compare performance in context.

Results stay attached to the evidence slice, model digest and hardware target that produced them.

Demonstration benchmarkScores below illustrate the Manyform evaluation model. They are not production claims about downloadable models.

IVC real holdout / EV-228

Industrial Visual Event Benchmark

Real-only protected observations spanning package, robot-cell and warehouse visual conditions.

1,248 real observations5 operational slicesZero training descendants

Leading real score94.2macro F1 · package states

RankModel releaseRealSyntheticFalse alarmsp95 / memoryTargetStatus

1Package State Detectorv0.8.393.695.11.8%27 ms1.2 GBJetson Orin NXEvaluated

2EdgeSight VLM Smallv1.4.191.892.42.4%142 ms3.1 GBQCS6490Released

3Cell Boundary VLMv0.5.089.493.23.1%618 ms9.4 GBAWS g5.xlargeRelease pending

Operational slices

A headline score never stands alone.

Inspect the conditions where a model needs more evidence before it is considered for a target.

All real holdout

1,248 observations

91.8

Package deformation

264 observations

94.2

Low light

218 observations

86.4

Partial occlusion

196 observations

88.7

Camera shift

172 observations

90.1

Release gate

Release belongs to a model-target pair.

Identity verifies the artifact. Fit checks the declared hardware and operating envelope. Reach confirms bounded export through a declared connector. A release record is not a safety certification.

1 Identity2 Fit3 Reach