The trust page

Promises, with receipts.

Most software promises are policies — words a company can quietly change. Slayt's are architecture: enforced by the database schema, by permission rules, and by tests that fail the build the day a promise would break. We maintain a written AI risk-management program mapped to the NIST AI Risk Management Framework; this page is the plain-English version of it.

What AI does here — and what it can never do.

Three uses, each fenced. Eligibility isn't on this list because it isn't AI at all — it's boolean logic over what talent declared, in plain SQL, with no model anywhere near it.

Tape QC

Flags audio level, framing, brightness, and length — advice shown to the actor, who can always submit anyway.

Can neverReject, block, or rank a submission.

Summarize & group

One observable, non-evaluative line per tape; groups similar reads so casting reviews coherent runs.

Can neverScore, grade, recommend, or comment on protected traits.

Brand match (future)

For creator campaigns: compares accounts and audiences to a brand's brief.

Can neverRank a person's suitability for employment.

The promises, and what enforces each one.

Every line here traces to a specific database rule or test in our codebase. If an enforcement fails, the build fails.

No automated rejectionsHuman decisions

There is no code path that lets software mark a person "pass." The refusal is enforced in two independent layers — the API guard and the database schema itself — each proven by its own tests.

No AI ranking of peopleOrganize, never decide

Model prompts are pinned by adversarial tests that feed them ranking-bait and fail if anything resembling a score, grade, or recommendation comes back. A prompt change that regresses the constraints cannot deploy.

Nothing inferred about youSelf-declared only

Matching uses only structured fields talent declared themselves. Undeclared means unmatchable — the system would rather show nothing than guess. ZIP code, school, name, and photos are never matching inputs.

Every decision on the recordImmutable audit log

Each status change writes an append-only event — who, when, why. The log physically refuses edits and deletes, even from administrators, and is retained for three years minimum.

Payment can't touch visibilityFree for talent, forever

There is no billing code in the product's data layer at all — and a canary test fails the entire build the day any appears without an isolation design. Sort order structurally cannot read payment state, because there is none to read.

Minors are guarded by the databaseNot by policy

Date of birth is encrypted at rest and used only to detect minors and compute a playable-age band. A minor's profile cannot exist without guardian contact — the database constraint refuses it.

Transparency without pressureAn ending, not a feed

Every submitter sees the role's honest status and real numbers for the life of their record. No read receipts, no streaks, no scores, no "submit more" — and casting's working notes never render to talent.

Your data is for castingNot for training

Tapes and profiles are never used to train models. Talent can export everything, free, before any retention purge — your work stays yours.

The sentence on every talent page.

"This tool helps casting organize submissions. All casting decisions are made by people."

That notice is a shared component rendered on every talent-facing surface — one copy in the codebase, so it can't quietly drift. The policy behind this page is reviewed quarterly, at every AI-touching launch, and whenever the relevant law moves (Texas HB 149, Illinois HB 3773, NYC Local Law 144, Colorado SB 24-205). A confirmed control failure is a stop-ship: the feature goes dark, the failure becomes a test, and the fix must pass it before re-enable.