# Egocentric Data Provider Evaluation Checklist

Version 1.0 · Published 24 July 2026 · EGXO Data

Use this checklist during technical, legal, privacy, and procurement review. Score only what the provider supports with a representative episode, a production record, a pinned term, or a reproducible acceptance test.

## Hard gates

A provider fails the pilot if any of these remain unresolved:

- The intended model-training or commercial use is not clearly permitted.
- Material consent, privacy, bystander, retention, or deletion requirements are not met.
- Critical actions are not visible from the proposed viewpoint.
- Required streams cannot be reconstructed and aligned on a common timeline.
- A representative release fails the buyer's preprocessing or training loader.
- Dataset, schema, calibration, and terms versions cannot be pinned.

## 100-point scorecard

### 1. Deployment and task fit — 20 points

- [ ] Target model, policy, perception task, or evaluation claim is explicit. (3)
- [ ] Task, object, tool, environment, actor, and outcome coverage matches deployment. (5)
- [ ] Hard cases and expected sources of variation are sampled deliberately. (4)
- [ ] Train, validation, test, and held-out slices match the generalization claim. (4)
- [ ] A representative training or evaluation record is defined before collection. (4)

Evidence:

Result:

Limitations and remediation:

### 2. Viewpoint and observability — 15 points

- [ ] Camera placement is justified against what the model must observe. (3)
- [ ] Hands, manipulated objects, contact transitions, and terminal state are measurable. (4)
- [ ] Occlusion, framing drift, blur, lighting, and off-screen events are reported. (4)
- [ ] Critical-action visibility and usable-take yield meet agreed thresholds. (4)

Evidence:

Result:

Limitations and remediation:

### 3. Time, calibration, and metadata integrity — 15 points

- [ ] Clock authority, timestamps, offsets, drift, gaps, and resampling are documented. (4)
- [ ] Intrinsics, distortion, extrinsics, axes, frames, units, and calibration version are pinned. (4)
- [ ] Every field is identified as measured, derived, annotated, or synthetic. (3)
- [ ] Missing streams, uncertainty, valid masks, and transformation history are explicit. (2)
- [ ] A complete episode reconstructs without undocumented filename assumptions. (2)

Evidence:

Result:

Limitations and remediation:

### 4. Provenance, consent, privacy, and rights — 20 points

- [ ] Contributor, device, collection, processing, review, and release provenance is traceable. (4)
- [ ] Consent, bystander, privacy, redaction, retention, deletion, and revocation controls are documented. (5)
- [ ] Model-training and commercial-use permission are explicit and versioned. (4)
- [ ] Redistribution, subcontractor access, derivatives, geography, and publication terms are separate. (4)
- [ ] Data rights are not inferred from a code or repository license. (3)

Evidence:

Result:

Limitations and remediation:

### 5. QA evidence and rejection trail — 15 points

- [ ] Acceptance tests were defined before collection. (3)
- [ ] Usable yield and rejection reasons are reported by task, device, and environment. (3)
- [ ] Accepted, rejected, repaired, and ambiguous examples are reviewable. (3)
- [ ] Annotation agreement, escalation, correction, and recurring failure rates are measured. (3)
- [ ] Privacy, rights, schema, synchronization, and loader failures are part of QA. (3)

Evidence:

Result:

Limitations and remediation:

### 6. Delivery and loader readiness — 15 points

- [ ] Dataset, schema, format, and loader versions are pinned. (3)
- [ ] Checksums, manifests, splits, episode boundaries, and release notes are complete. (3)
- [ ] Media decode, modality alignment, masks, and batch sampling pass automatically. (4)
- [ ] A representative release passes the buyer's real preprocessing or training loader. (3)
- [ ] Manual repair time, issue ownership, and support response are documented. (2)

Evidence:

Result:

Limitations and remediation:

## Decision record

- Provider:
- Pilot scope:
- Dataset or release version:
- Terms reviewed and date:
- Total score:
- Hard-gate result:
- Approved use:
- Open risks:
- Required remediation:
- Decision owner:
- Review date:

The score supports a decision; it does not replace legal, privacy, security, or model-specific review.

Source guide: https://egxodata.com/resources/how-to-evaluate-egocentric-data-provider