Labels you can audit: benchmark-adopted perception, mask-backed boxes, and an agent that runs the pipeline
Two weeks ago we published our first perception number: precision 0.449, recall 0.298 against the public BDD100K ground truth, zero-shot, no fine-tuning. This post is about what happened when we treated that number as a starting line instead of a press release.
Adopt detectors by measurement, not reputation
We benchmarked four open-vocabulary detectors on the same 284 ground-truth images with a fixed canonical vocabulary, fitting per-class confidence floors on half the set and scoring on the held-out half. The default everyone reaches for did not win.
| Detector | Holdout P | Holdout R |
|---|---|---|
| OWLv2-base ensemble | 0.723 | 0.567 |
| Grounding-DINO base | 0.550 | 0.564 |
| SAM3 (adopted July 2026) | 0.770 | 0.694 |
| Grounding-DINO tiny | 0.450 | 0.487 |
| MM-Grounding-DINO tiny | 0.435 | 0.475 |
Update, July 2026: we re-ran the identical protocol with SAM3, the first Segment Anything release that takes concept prompts and returns boxes and masks from one model. It beat our adopted OWLv2 path on both axes (precision 0.770 vs 0.723, recall 0.694 vs 0.567) on the held-out half, so SAM3 is now the default detector. Every earlier number in this post stays as published; the table above gains one row.
The shipped product path (adopted detector, vocabulary-dense prompting, per-class floors, cross-class NMS, and readiness QA) now measures precision 0.542 / recall 0.466 on the same ground truth. That is +9 points of precision and +17 points of recall over our published baseline, on the honest, shipped output. The whole harness is drop-in: any detector that can emit boxes gets the same scorecard.
Boxes with boundaries: Grounded-SAM everywhere
The detector names the object; SAM draws its boundary. Every grounded label Euler ships is now refined to its mask's tight bounding rectangle and carries the mask evidence with it, so our label QA can demand real segmentation support. In our driving demo project, 98% of labels are mask-confirmed. When a label is uncertain, it does not silently vanish: it ships flagged for human review, with a workbench that filters by category, draws and renames boxes, records controlled attributes like color, occlusion, and state, using your ontology per use-case, and folds every human edit into the sealed export.
Video, across time
The same machinery now works on video: keyframe detection with the adopted detector, SAM boundary propagation between keyframes, and a timeline editor where tracks follow the playing video, sub-task spans sit on a scrubbable bar, and every track and span is editable. Sub-task segmentation runs on timestamped frame sheets with Gemini 3.5 Flash, the approach the field has converged on, and lands as editable spans instead of opaque metadata. A public-benchmark comparison on WGO-Bench is next; we will publish our segmentation F1 and cost per video hour against the numbers already out there.
An agent that operates the platform
All of this is drivable from a conversation. The Agent Center plans each chat turn with an LLM that is constrained to Euler's typed, audited operations: create projects, ingest buckets, analyze data-to-policy fit, recommend a training target with readiness evidence, run pipelines, curate, seal exports. It is grounded in the live state of your workspace, and paid steps stay gated behind your approval. Ask it to set up a drone-flight project and run readiness on real FPV racing trajectories; we did, and the kinematic checks flagged every laser-tracker glitch in the raw data without being told what to look for.
Every model, accountable
Which model serves which surface (embedding, detection, segmentation, annotation, agent planning) is now a registry with the benchmark evidence attached, queryable from the API. Swapping a model for one use-case is a data change that has to bring its numbers with it. That is the whole thesis: data readiness you can audit, from the first ingest to the last label.