Previz/Legal/Trust & Content Security

    Built to be
    the sanctioned zone.

    Studios now publish rules for how their partners may use generative tools. Previz — temporary, non-final, pre-production planning — is exactly the use those rules sanction. This page shows how the platform maps to them, and where our own compliance work stands, in the open.

    Frames made in Previz. Every one of them can state which engine produced it, from which inputs — which is what the rest of this page is about.

    The commitments

    01

    Never trained on

    Your scripts, frames, and sequences are never used to train our models or anyone else's. Inference-only vendor contracts, one server-side chokepoint, named-vendor exclusions enforced by the router.

    The enforcement mechanics
    02

    Temporary by design

    This is a previz platform. Its outputs are planning artifacts — boards, animatics, camera tests — built to be replaced by real photography. The export path (EDL, FCPXML, PDF boards) hands the plan to your cutting room; it does not compete with it.

    03

    Provenance on every asset

    Lineage records which engine produced every frame, from which inputs, automatically — a per-asset birth certificate you can show a rights holder, a client, or a compliance review.

    The Partner Principles

    Five rules the studios wrote. How we hold each one.

    In 2025 the major streamers published guiding principles for generative AI use by their production partners. Paraphrased below — we claim no affiliation or endorsement — with, against each one, what the platform actually does. Every row is verifiable in the product.

    01

    Generated material should not replicate identifiable elements of works you don't own.

    Generation anchors to YOUR references: continuity locks pin your characters, props, and locations, and the drift loop re-checks every reappearance against them. Per-asset lineage shows exactly which inputs shaped which frame, so an IP review is a lookup, not an investigation.

    02

    Tools must not store, reuse, or train on production inputs or outputs.

    The zero-training-retention architecture: commercial inference endpoints only, no consumer tiers, process-and-return vendor terms, server-side keys, and your work persisting only in your tenant-scoped storage.

    03

    Use enterprise-secured environments, not personal accounts.

    Row-level security on every table, tenant- and project-scoped storage paths, org accounts, and per-project vendor and jurisdiction exclusions enforced server-side before anything is generated. Private deployment — dedicated tenant, VPC, air-gapped — is the funded roadmap for engagements where the answer must be 'no third party at all'; the platform running today is the shared one.

    04

    Outputs are temporary working material, not final deliverables.

    The product is previz: storyboards, animatics with scratch audio, camera and lens exploration. Final-frame delivery deliberately exits to your NLE via EDL/FCPXML — the platform's own renders are labeled animatics.

    05

    Performances and union-covered work are not replaced without consent — when in doubt, escalate.

    Scratch dialogue is labeled scratch and exports as comments, not audio stems. Voice cloning ships behind a consent gate: a clone is refused, before any charge, unless it is made against a consent record for that voice — one naming whose voice it is and, for another person's voice, linking the release they signed — and a cloned voice stops performing if that record lapses. There is no face-swap, and no tooling that builds a model of a performer's face or body. Provenance gives productions the disclosure trail their agreements require, and anything beyond these guardrails belongs in your approval chain — not silently in a render.

    Rights & Consent

    Whose face, and who said yes.

    Continuity means holding a person's likeness steady across hundreds of frames, which makes permission a product problem rather than a paperwork one. Here is exactly where that stands. The rows marked planned are not shipped, and we would rather you know that from us than discover it in a rights review.

    Whose referencesIdentity locks anchor only to reference material the customer uploads into their own tenant. The platform ships no library of real people to generate fromLive in product
    Digital replicasVoice cloning ships, and only against a consent record for that voice, checked when the clone is made and before any charge: it must be your account's, name whose voice it is, say whether it is your own, another person's or a synthetic character, link the release they signed if it is another person's, and be neither revoked nor past its end date. One record backs one voice, and a cloned voice is refused whenever its record has lapsed. The record is your production's statement: we check it is complete, not that the release is genuine, and revoking a record does not yet stop a voice already made from it. There is no face-swap and no tooling that builds a model of a performer's face or body; lip-sync re-times the mouth in a video you supply to audio you supply and has no consent gate of its own. Voice work is labeled scratch and exports as comments rather than audio stemsLive, with a limit
    Disclosure trailPer-asset lineage — engine, inputs, edit ancestry — recorded at creation, exportable as the record a rights holder or a broadcaster asks forLive in product
    Vendor exclusionsIf your legal team excludes a provider, the router enforces it per tenant. The exclusion is a routing rule, not a promiseLive in product
    No training, everInference-only vendor terms with no consumer tiers, so nothing generated on a production becomes training data anywhere downstreamLive in product
    Consent recordsStructured grants bound to the identity lock — who permitted what, for which use, until when — inherited by every asset generated from that lockPlanned
    Expiry enforcementGeneration blocked at request time when a grant has lapsed or does not cover the intended use, defaulting to blocked rather than allowedPlanned
    Clearance packA per-project export pairing every likeness in the cut with the grant that permitted it, for the approval chain and for disclosure obligationsPlanned

    One thing worth saying plainly: the platform ships no library of real people to generate from, and it has no tooling for building a replica of a performer's face or body. The one replica it can make is a voice, and only against a consent record your production writes — a record of what you tell us, not a verification of it. Every identity the system holds steady got there because someone on your production put it there.

    Made with Previz

    Frames made in Previz. Each carries the lineage record described above — which engine, from which inputs.

    Content-Security Self-Assessment

    Where we stand, control by control.

    Our self-assessment against the MPA content-security best practices — the framework behind Trusted Partner Network assessments — published here rather than kept in a drawer. Statuses are conservative: "Live in product" means enforced in the running platform today, and we say "Planned" where the work has not shipped. No certification is claimed before it exists.

    Content isolationDatabase-enforced row-level security on every table, and tenant/project-scoped storage paths. Where a bucket is public, those policies govern authenticated access rather than the public read pathLive, with a limit
    Access controlPer-user authentication, org roles, and least-privilege service keys held server-side onlyLive in product
    Asset provenanceAutomatic per-asset lineage: engine, inputs, and edit ancestry recorded at creationLive in product
    Transfer securityTLS in transit everywhere; tenant- and project-scoped object paths. Signed, expiring delivery is a dedicated-deployment property, not a shared-platform oneLive, with a limit
    Generation auditPer-generation usage records (who, what, which engine, when) retained per tenantLive in product
    DeletionA deletion request removes database records and every stored object under the tenant's paths, reporting the object count erased; backups age out within 30 daysLive in product
    Incident responseWritten IR policy with 24-hour customer notification commitment, published in the policy packPolicy in place
    Secure developmentDocumented secure-development policy; RLS required in the same migration as any new tablePolicy in place
    TPN assessmentTrusted Partner Network membership and Blue Shield self-report submissionPlanned
    SOC 2SOC 2 Type I engagementPlanned
    Org audit logOrg-level administrative audit log surfaced to enterprise adminsPlanned
    SSOSAML / OIDC single sign-on for enterprise tenantsPlanned

    The full policy set behind the "Policy in place" rows — access control, data protection & deletion, incident response, secure development, vendor management, and the rest — ships as a versioned documentation pack, published and readable now rather than held behind an NDA.

    Bring your compliance team. We like it that way.

    Every claim on this page is either visible in the product — lineage, isolation, exports — or documented in the policy pack, which is published rather than gated. Security questionnaires get answered directly, with engineers in the room.

    Plain language

    This page is our own statement of practice — an honest self-assessment, not a contract. The legal centre carries a plain-language summary of it. The binding documents are the Terms, Privacy Policy, and Acceptable Use.