A readiness certificate you can inspect
What Euler measures across multimodal robot data, and why an evidence-first verdict deserves more trust than a single readiness score.
A dataset can look healthy in a browser and still hide a clock fault, a mismatched companion stream, an unreliable caption, or a narrow distribution. Euler Data Certification finds those problems before training and explains each judgment with a receipt.
The output is an Euler Certified evidence bundle, not a rewritten training dataset. It records what was observed, which policy was applied, which checks were applicable, how confident each result was, and why the dataset received its verdict. The customer's raw frames are not copied into the bundle.
Four verdicts, because “pass/fail” is not honest enough
This distinction matters operationally. An annotation that can be corrected should not condemn a valid recording, while a duplicated clock value should not be waved through as a review note. Coverage is reported separately from quality. Missing evidence cannot hide inside an average.
The five signal families
The certificate names the evidence it checks without publishing a recipe for gaming the result. Euler reports five signal families.
Timeline & Sync
Clock integrity comes first: monotonic sensor clocks, gaps, and cross-stream alignment. For video and IMU, Euler derives a sub-frame alignment signal from the content itself. The report shows the residual and whether the alignment signal is unique. In a recent pilot, keyframe residuals reached sub-millisecond precision; when the correlation peak was ambiguous, the certificate marked it low confidence.
Kinematics & Physics
Smoothness and frequency-domain sanity checks look for motion that is erratic, aliased, or physically implausible. The goal is not to impose one style of motion on every robot. It is to flag recordings whose inertial signals do not describe a usable physical trajectory.
Visual Integrity
Sampled sharpness, exposure, contrast, calibration, and geometry checks show whether the visual stream can support its intended task. Euler uses one consistent measurement for each visual property across a receipt. Blur is not judged two different ways in the same report.
Task & Language / Annotation
Euler reports caption coverage and caption-to-video agreement using a deterministic two-sample self-consistency pass, guarding against a single sampled description becoming ground truth. When a source carries hand motion-capture annotations, a separate check compares the customer's own hand pose with the pixels. It judges the supplied annotation and never replaces it.
Dataset Distribution
Recording length and embedding diversity describe the set as a whole. They expose collections dominated by one motion, scene, or duration before that imbalance becomes model behavior.
Why the certificate is reproducible
The policy is part of the evidence
Every certificate pins a content-hashed policy. If the policy changes, the next report gets a new version; a previous certificate is never rewritten in place.
Applicability comes before judgment
Egocentric, teleop, drive-log, aerial, state-only, and vision-asset data do not share one contract. Euler activates only the metrics the dataset can honestly support.
Confidence travels with the result
Observed values, inputs, provenance, confidence, and the applicability decision stay attached to each metric. The result never leaves its evidence behind.
Absence is not success
Missing modalities reduce evidence coverage. They do not become a passing value, and they do not become an arbitrary zero.
The sealed bundle includes a human-readable report, a machine-readable report, per-metric receipts, diagnostics, the pinned policy, and a README. The customer can reproduce and audit the decision without creating another copy of the training corpus.
What recent pilots surfaced
In a recent egocentric multimodal-data pilot, Euler audited a partner's 14-recording dataset against a versioned policy of roughly 28 rules. That produced 392 per-recording metric assessments with 100% required-evidence coverage. Thirteen recordings were training-usable.
The remaining recording had two specific faults: a duplicated sensor timestamp and a 15-frame companion-video mismatch. Each fault arrived with a receipt naming the offending stream. Euler surfaced both before a single training step.
Caption-to-video agreement landed between 0.95 and 1.0 across those recordings under the two-sample self-consistency pass. In a separate annotation pilot, Euler paired per-frame hand pose with SOP-conformant dense text and used agreement across independent samples as a hallucination guard. Both pilots followed the same principle: strong evidence first, automation second.
What Euler Certified does not claim
Certification does not promise that one dataset guarantees a good model, and it does not turn an inapplicable measurement into a universal standard. It establishes that a declared data contract was assessed with a pinned policy, that every required result has inspectable evidence, and that uncertainty stayed visible all the way to the verdict.
That is the useful boundary: do not trust a score on faith. Inspect the receipts, reproduce the decision, and see its limits.
Run it through Euler Data Certification.