Browse documentationData and privacy
Reviewers receive only the data needed for their task
GenMedia separates evaluation access, task metadata, results, and reviewer operations. The interface hides candidate identity during review, limits external accounts to explicit assignments, and keeps identity-linked reliability data out of ordinary result and export paths.
What a task stores
- Evaluation Task setup
Project, task protocol, criteria, audience, presentation, and comparison design.
- Generation lineage
Prompts, seeds, reference inputs, endpoint labels, request settings, runs, and candidate media.
- Review record
Presented order, choice or rating, tie or skip state, permitted comments, and session quality signals.
- Result evidence
Observed counts, eligible records, aggregate statistics, and role-level aggregate summaries.
What reviewers see
An internal evaluator receives the Task-pinned prompt copy, task instructions, criteria, shared input context, required source-conditioning media, permitted optional references, and blind candidates. Source media and candidates use assignment-scoped opaque relay URLs. Model identity, private object locations, and aggregate results remain gated until the role and completion rules allow them.
An external collaborator receives less: a neutral assignment, the minimum evaluation context, and the blind candidates. Projects, task status, model labels, aggregate progress, results, and exports remain hidden.
Lineage and identity
Once review begins, the product protects inputs and candidate identity from changes that would reinterpret an earlier vote. Normal exports remove stable reviewer identity. Separate administrative access is required for reviewer-linked audit data.
Safe data handling
Do not place secrets, personal data, customer-confidential material, or unlicensed media in prompts, comments, task titles, imports, or shared DeltaFrame sessions. Use the smallest audience that can answer the evaluation question and share exports only through an approved channel.