Data Assessment

Use the data-assessment workspace (formerly data-certification; the old id keeps resolving for existing workspaces) when the deliverable is an auditable readiness report rather than a transformed training dataset. The workflow is generic across robotics programs and keeps policy choices separate from source measurements.

What Euler measures

For supported multimodal HDF5 recordings, Euler preserves video-rate state and native-rate IMU streams separately and records:

• source-clock rate, gaps, and monotonicity;
• direct video-to-IMU keyframe residuals;
• content-derived sub-frame lag from visual-SLAM angular speed and IMU gyroscope magnitude, with peak correlation and confidence;
• companion-video frame counts;
• sampled exposure, temporal difference, and blur evidence;
• Welch power spectral density ratios and Allan deviation when duration permits;
• calibration inventory, stereo baseline, transform orthogonality, depth alignment, and caption structure;
• annotation quality of the source’s own captions: sub-task timeline coverage, dense action-label coverage, boundary sanity (in-bounds, ordered windows), and caption-video agreement.

Annotation quality

When a recording carries hierarchical captions, Euler audits them instead of merely parsing them. Coverage and boundary sanity are deterministic measurements from the caption windows themselves. Caption-video agreement is a run-time audit: timestamped contact-sheet frames sampled from the recording and the annotation windows go to the vision model, every window receives a match / mismatch / unclear verdict with a short note, and the per-recording agreement ratio becomes the language.caption_video_agreement metric. Mismatched windows are kept as receipts (window, label, reason) in the report diagnostics. When the model key or GPU lane is absent the metric reads not_assessable with its reason; it never silently passes. Euler measures the annotations; it does not rewrite them.

The content-derived lag is an independent motion-signal diagnostic, not a replacement for hardware clock evidence. A broad or low-excitation correlation peak remains review_required even when its numeric lag is within the ideal boundary. Stereo reprojection stays not_assessable unless the source includes trusted correspondences or target observations.

Create the workspace

Create a workspace with use case data-assessment, then create one project per source population or collection contract. Run ingest normally. Export Center keeps the standard layout: the Target readiness panel validates the assessed episodes against every target profile exactly as in any other workspace, and a Readiness certificate tab holds the sealed report bundles. Successful runs seal report bundles, never a training dataset, for this use case.

Review or version a policy

GET /v1/projects/{project_id}/certification/policies returns the immutable policy history. A reviewer can change target, ideal boundaries, or review boundaries with POST /v1/projects/{project_id}/certification/policies; the server assigns the next version and records the policy hash in the audit log.

Rate and quality thresholds are target-dependent. Keep exact integrity checks strict by default, and change a boundary only when the downstream training contract supports it. For example, accepting one terminal video frame must be an explicit policy decision rather than an ingest-side correction.

Inspect and export

GET /v1/projects/{project_id}/certification/report?policy_version=1 returns the machine-readable report. POST /v1/projects/{project_id}/certification/export?policy_version=1 seals an Euler Certified bundle into Export Center containing:

• report.json: canonical result and stable certification ID;
• report.html: shareable summary;
• metric-results.csv: every episode and metric decision;
• diagnostics.json: selected derived clock, frame, visual, spectral, calibration, depth, and caption evidence without raw media;
• policy.json: exact thresholds, formulas, and policy hash;
• README.md: bundle interpretation.

The archive manifest sets contains_training_data to false.

Privacy-constrained local audits

An administrator can run the same open-source engine beside private data, then send the derived report.json to POST /v1/projects/{project_id}/certification/import. The report project ID must match and its policy hash must already be registered on the destination project. Export Center marks the manifest origin as local_report_import; the report schema contains evidence and decisions, not media.

Interpreting the verdict

• certified: every required metric meets the pinned ideal boundary.
• review_required: no required hard failure exists, but evidence is missing, borderline, or low-confidence.
• not_ready: at least one required metric crosses its review boundary.
• not_assessable: no required evidence was available.

Coverage is reported alongside the verdict. It is the fraction of required episode-metric decisions that had applicable evidence, never a substitute for the verdict or a model-performance guarantee.