The department

Four desks, one producer's answer

A studio doesn't ask one person "is this script risky?" — it runs four independent departments that never talk to each other, then reconciles them. ScriptRisk mirrors that, desk for desk.

CC

Clearance Counsel

Rights & clearances — the desk that keeps you out of court.

  • Music — two licenses, sync + master, different owners. Both chains chased.
  • Brands — a visible logo is fine; a mocked or blamed one is disparagement. Only the second flags.
  • Real people — right of publicity, and the defamation line: documented record vs. invention.
  • Created works — artwork, tattoos, film clips, with the 95-year public-domain arithmetic.
  • Venues — permit vs. agreement vs. "that casino will never say yes — plan a set."
sync_licensemaster_useright_of_publicitydefamationtrademarktrade_libel_venueartwork
RB

Ratings Board

What will the MPA actually rate this — and what would move it?

  • Counts, never estimates — every F-bomb cited by scene; violence, drugs, sexuality, themes.
  • Evidence, not opinion — your profile against 4,733 official CARA rating rationales, plus the measured CARA boundary (90% per-group coverage, 92% aggregate).
  • Cuts re-run live — check a cut in the What-If simulator and the comparables recompute.
rating_languagerating_violencerating_drug_userating_sexualityprediction + comparables
SU

Safety Underwriter

The line producer's early-warning system for physical risk.

  • Everything that can hurt someone — stunts, pyro, water, vehicles, firearms, animals, minors — CSATF bulletins cited.
  • Remedies read like a shooting plan — which specialist, what it costs, what it adds to the schedule.
  • Gags get the method — rig, dummy, VFX takeover — never a prohibition.
stunt_*fire_safetyminor_safetyanimal_safetyfirearms_blanks
TC

Territory Censor

What gets cut or banned, market by market — US · UK · CN · UAE.

  • Geo-targeted research — the CN questions are asked from China's web, the UAE's from theirs.
  • Depicted vs. recounted — a ghost story in dialogue is not a ghost on screen; story-critical vs. an easy localized cut.
  • Six-axis closing sweep per strict territory — supernatural, drugs, alcohol, religion, sexuality, state authority.
territory_cn_*territory_uae_*territory_uk_*territory_us_*
Under the hood

The architecture, honestly

Agents decide what to investigate; tools decide what's allowed to be true. Every box below is real and running — nothing in this diagram is aspirational.

Your scriptparsed deterministically — no model
→
Triageevery entity onto a desk's worklist
→
Four desks in parallelclearance splits into batch agents on big scripts — bounded context by construction
→
Blinded verifiersees only claim + citation — rejects what the source doesn't support
→
Adjudicatormerges duplicates, resolves conflicting remedies
→
Your reportevery finding cited, priced, fixed
Parallel Search · Extract · Task APIthe desks' only window to the world — live web retrieval with verbatim excerpts
ClickHouse · 4,733 CARA rationalesthe ratings evidence — films whose official rationales sit nearest yours, each with a source URL
Provenance registryevery retrieved excerpt, logged — the citation gate checks against THIS, not the model's memory
Why you can trust it

Accuracy is enforced by code, not requested by prompt

1

No citation, no finding

The filing tool structurally rejects any flag without a citation. A finding the model "just knows" cannot exist in your report.

2

Quotes must be real

Every cited excerpt is checked character-for-character against what was actually retrieved this run. Paraphrase or invent, and the filing bounces.

3

Blinded cross-examination

An independent verifier re-reads each citation without the desk's reasoning. Rejected findings never touch the score — and they're shown, struck through, with the reason.

4

Numbers come from data

The rating prediction is a nearest-neighbor search over released films, not a model's opinion — and each comparable carries its own source.

5

Honest unknowns

What can't be verified becomes an open question, not a confident guess. "We couldn't prove it" appears in your report exactly that way.