Technical overview

How persistent identity works

Version 1.0 · October 2026 · Ananya Das

Summary: Pulse AI treats the person, not the video, as the unit of the system. A sender is enrolled once from consented source material. Every message is generated from that identity, checked against it, and approved before it sends. Consent and revocation are attached to the identity, so they apply everywhere it is used. Each section below marks what is Live today, what is Next, and what is Later.

1. The problem, stated precisely

Generation models render a person on demand. Each output is produced independently, and quality checks score each video on its own: realism, lip-sync, artifacts and, at most, identity similarity to the source. Nothing holds constant the property a recipient actually relies on across a sequence: that the person in message three is the same person as in message one, and that they authorized the message.

We call the failure identity drift: the change in a person's perceived identity across AI-generated outputs that are meant to depict the same person. In a sales sequence it is costly because it is cumulative. Every video can pass a per-video check while the sequence as a whole stops looking like one sender.

The remedy is a persistent identity: one record of a real person, created once with their consent, that every AI-generated message depicting them is bound to. It is enrolled once, bound to consent, held consistent against one reference, and revocable everywhere. The concept and its requirements are set out in our position paper [2]; the measurement in our working paper [1].

2. The identity model

Everything in the system hangs off one record per person.

Identity one per person · active or revokedConsent recordsigned release · approved usesReference material consented source · versioned Campaigns reuse the identity, never re-upload Outputs checked · approved · logged

ObjectWhat it holdsStatus

IdentityOne record per person, with a status: active or revoked. The unit everything else attaches to.Next as a standalone, reusable record. Today the enrolled sender exists per deployment.

Consent recordThe signed release, the uses it approves, and the date. Required before any generation.Live

Reference materialThe consented source the identity is generated from and checked against. Versioned, so later outputs are compared with the same reference.Live (single reference) · Next (versioning)

CampaignsSequences that use an identity. A new campaign selects an enrolled identity instead of uploading new material.Next

OutputsEach generated message, linked to the identity and reference it came from, with its review decision.Live (human review) · Next (stored consistency history)

3. From consent to delivery

ConsentEnrollGenerateCheckApproveDeliverRevoke at any point: generation and delivery stop for every use of the identity

  1. Consent. The sender receives written notice and signs a release naming the approved uses. Nothing is generated before this. Live

  2. Enroll. Approved source material is enrolled as the sender's reference identity. Live

  3. Generate. Each message in the sequence is generated from the reference identity and an approved script. Live

  4. Check. Each output is compared with the reference identity before it can be sent. Today this is a human review using the published evaluation method; automated scoring follows validation. Live (human review) · Next (automated)

  5. Approve. The customer approves every message. Nothing sends automatically. Live

  6. Deliver. Approved messages are delivered by email with a hosted video link. Live

4. The generation pipeline

Generation is a multi-stage pipeline. The identity check sits after generation and before anything can reach the approval queue.

Stage

Purpose

Lip-sync generationRenders the enrolled sender delivering the approved script.

UpscalingRestores detail for a natural, high-resolution result.

Frame interpolationSmooths motion between frames.

Colour gradingKeeps tone consistent across every message in the sequence.

Identity checkCompares the output with the reference identity. Outputs that fail never reach the send queue.

The pipeline is model-independent by design: the identity and its checks sit outside any single generation model, so models can change without the identity changing.

5. Measuring persistence

Our working paper, Identity Persistence in AI-Generated Video: A Sequence Evaluation Framework for Founder-Led Outbound (Zenodo, CC BY 4.0) [1], introduces the Sequence Persistence Score (SPS): a way to evaluate whether one identity stays credible as the same sender across an AI-generated video sequence.

  • Reference-based. Every output is judged against the enrolled reference, not only against the previous frame or video, so drift that builds slowly across a sequence still registers.

  • Sequence-level. The unit of evaluation is the sequence (first touch, follow-up, after the call), not a single video.

  • Open. The evaluation code is public under the MIT licence: github.com/byananya/sequence-persistence-evaluation.

  • Status. A working paper, not yet peer reviewed. A broader framework covering face, voice and motion, calibrated against human review, is future work.

Read the paper for the method in full. This page summarizes it; it adds nothing beyond what is published.

6. Requirements, and where Pulse AI stands

The position paper [2] sets out the requirements any persistent identity infrastructure must meet. This is how Pulse AI meets each one today.

Requirement: How Pulse AI meets it

Status

Consent before generationSigned release before any generation; outputs stay within approved uses

Live

One fixed referenceEvery output judged against the enrolled referenceLive (single reference) · Next (versioning)Sequence-level judgmentConsistency assessed across the whole sequenceLive (human review) · Next (automated)Human approvalThe customer approves every message before it sendsLiveRevocation everywhereWithdrawal stops generation and delivery and starts deletionLive (on request) · Next (automated)Model independenceThe identity and its checks sit outside any single generation modelLive (by design)VerifiabilityAnyone receiving a message can check it was authorizedNext (signed record) · Later (verification page)

7. What it protects against

RiskWhat could go wrongHow Pulse AI handles itStatusIdentity driftThe sender gradually stops looking like themselves across a sequenceEvery output checked against one reference identity before sendingLive (human review)Generation without consentSomeone's likeness is generated without their permissionNo signed release, no generation. The customer confirms the sender is the person in the materialLiveUse beyond approvalA likeness is used for a message the person never approvedThe release names approved uses; the customer approves every messageLiveStale authorizationGeneration continues after the person withdraws consentRevocation stops generation and delivery and starts deletion. Today handled on request; automated across every use nextLive (on request) · Next (automated)Altered or replayed videoA delivered video is edited, or reused out of contextA signed record for each approved output, so any change is detectableNextRecipient doubtA recipient can't tell whether the person authorized the messageA verification page for each message, showing authorization and current statusLater

8. Data handling

  • Source material and face data are used only for the approved uses, never sold, and never used to train models.

  • Retention and destruction follow a published schedule; deletion on request.

  • Face data is biometric data under several US state laws (Illinois, Texas, Washington, Colorado) and is handled as such.

Full detail: Privacy and Biometric Data Policy.

9. Roadmap

StageWhat it addsLiveConsented enrollment, sequence generation from one sender, human-reviewed identity check, approval on every message, email deliveryNextA reusable identity across campaigns, versioned reference material, stored consistency history per identity, automated revocation, a signed record for each approved outputLaterIdentity owned by the person and lent to the companies they work with; verification by anyone who receives a message

10. References

  1. Das, A. (2026). Identity Persistence in AI-Generated Video: A Sequence Evaluation Framework for Founder-Led Outbound. Working paper, Zenodo. doi.org/10.5281/zenodo.21620531

Questions: ananya@pulseai.pro

Home Technical overviewPrivacy & biometric dataTermsceo@pulseai.pro


© 2026 Pulse AI. Operated by Ananya Das.