Write the Start Frame End Frame AI endpoint contract

Treat the start image as the opening state and the end image as a separate destination. Record which subject, layout, lighting, and camera properties must remain stable between them. Inventory source rights for both images and mark any identity, product, text, or location detail that cannot change. Endpoint control is meaningful only when the transition has a defined editorial job and the two supplied states are compatible enough to review. In the decision log, preserve the reference set, and ask the claims reviewer to record human approval before the workflow transfer.

Inspect the path, not only the two supplied images

Choose boundary frames with one deliberate change, such as a door opening while camera position stays fixed, then inspect the entire transition rather than only the endpoints. Review intermediate frames at a consistent cadence for geometry, identity, contact, lighting, camera axis, and causal order. A model can reproduce the beginning and ending while inventing an unusable bridge, so acceptance belongs to the whole shot rather than two thumbnails. In the decision log, preserve the decision history, and ask the brand reviewer to record format readiness before the rights check.

• Opening-state fidelity In the decision log, preserve the decision history, and ask the accessibility reviewer to record format readiness before the scope confirmation.
• Causal intermediate motion In the decision log, preserve the handoff draft, and ask the rights reviewer to record reversal cost before the scope confirmation.
• Closing-state and edit usability In the decision log, preserve the input snapshot, and ask the workflow owner to record source fidelity before the final sign-off.

Separate capability, availability, and entitlement

Reject paths that teleport objects, reverse screen direction, invent a third composition, or satisfy the final frame by hiding discontinuity in the middle. Verify the exact vendor, model version, region, account surface, input constraints, and check date before relying on the feature. The documented Veo example proves a named control in that context; it does not prove current SEELE access or any free, downloadable, permanent, or open-ended use. In the decision log, preserve the review copy, and ask the source custodian to record control availability before the scope confirmation.

Keep the Start Frame End Frame AI evidence ledger narrow

Google DeepMind first-party documentation verifies a current first-and-last-frame control for Veo 3.1. Record the source owner, product or model name, verification date, surface or region, exact claim, and unresolved gap beside the test. Taxonomy eligibility and search demand explain why the page exists, but neither can be reused as proof that SEELE or another provider exposes the entire query as a current capability. For a controlled test, preserve the delivery checklist, and ask the workflow owner to record source fidelity before the evidence refresh.

• Google DeepMind lists first and last frame among the documented creative controls for Veo 3.1. For a controlled test, preserve the evidence table, and ask the continuity editor to record disclosure clarity before the delivery pass.