Identity Persistence in AI-Generated Video: A Sequence Evaluation Framework for Founder-Led Outbound

AI video tools are built and benchmarked to produce one good clip. Founder-led outbound does not sell in one clip — it sells across a sequence, and the sender is the trust signal carrying it. When the face or visual style shifts between the first outreach and the follow-up, the recipient stops believing the message came from a person and stops replying; the commercial cost of that shift is unmeasured because no existing benchmark looks across outputs.

This paper defines that failure mode as identity drift and specifies an evaluation framework that treats an outbound sequence, not a single video, as the unit of analysis. Three metrics are defined: face persistence, style persistence, and a composite Sequence Persistence Score (SPS). A sequence passes only if the sender remains recognisably the same person across every moment in it.

Face embeddings are extracted with InsightFace buffalo_l (ArcFace R100); the paper specifies frame sampling, pair counting, and the SPS weighting in full, with a worked numeric example. The generation pipeline is proprietary; the evaluation protocol is published in full so that any comparable pipeline can be scored against the same thresholds.

This is a working paper reporting a single-subject feasibility study. The face-persistence threshold is an uncalibrated acceptance floor, the style-persistence extractor is not finalised, and no outcome data has been collected. These gaps are stated explicitly and scoped as the next version's work.

Accompanying evaluation code: https://github.com/byananya/sequence-persistence-evaluation