🟢 GREENLIGHT

Automated pre-clearance for screenplays. Catch the legal landmines in a script while they are still free to fix.

---

The problem nobody codes for

Before a film can be shot, the production must obtain a script clearance report. Without it there is no E&O insurance, and without insurance no distributor will touch the film. It is not optional.

A human reads the screenplay and catalogues every named entity — character names, businesses, brands, phone numbers, addresses, licence plates, songs, artworks, hospitals, publications — then researches each one to determine whether it exists in the real world and whether using it invites a lawsuit.

That takes roughly a week and costs several thousand dollars per script.

And it happens too late. Clearance runs on the locked script. By then, renaming a bar means rebuilding set dressing, remaking props, and re-recording ADR. So productions negotiate, take risks, or pay for a licence.

The idea: shift clearance left

Software security moved from the end-of-cycle pentest to a linter in the editor. Screenplay clearance never made that move.

GREENLIGHT reads a screenplay, extracts every entity, verifies each one against the live web with traceable sources, and returns a clearance report in minutes — fast enough and cheap enough to run on every draft, not once at the end.

A writer who learns on page 12 that The Black Cat Tavern is a real, operating business fixes it in three seconds. The same fix six months later costs a shooting day.

---

Why this is not "an LLM with web search"

Depiction context is the risk multiplier. Naming a real bar is harmless. The same bar where a character deals drugs is defamation exposure. GREENLIGHT assigns a different verdict to the same entity depending on what the scene does with it — a judgement that requires reading the scene, not matching a string.

Replacements are re-verified. When an entity must change, the system generates an alternative, searches for it in turn, and only proposes it once it returns nothing real. A suggestion that has not been cleared is not a suggestion.

Diff mode makes per-draft runs viable. Draft v2 against v1: only changed entities are re-researched. Second pass in seconds instead of minutes. This is what turns clearance into CI.

Deterministic pre-verdicts cost nothing. A phone number in the 555-0100–555-0199 range is reserved for fiction by the North American Numbering Plan — CLEAR by rule, no network call. Same for RFC 2606 domains and government agencies named neutrally. This removes 10–15 % of entities from the queue before a single request is billed.

---

Pipeline

#	Phase	Engine	Deterministic
1	Ingest — Fountain / FDX → structured scenes	code	✅
2	Extract — typed entities + depiction context	Gemini Flash	schema-locked
3	Canonicalize — dedupe, alias resolution	code	✅
4	Research — web fan-out, risk-routed	Parallel Search	✅
5	Classify — verdict + rationale + citations	Gemini Pro	schema-locked
6	Suggest — generate, then re-verify replacement	Gemini + Parallel	—
7	Report — clearance report + margin annotations	code	✅
8	Diff — re-clear only the delta between drafts	code	✅

Determinism is enforced end to end: temperature = 0, strict responseSchema on every structured output, phases 1/3/7/8 contain no model call at all, and every finding records the prompt_version that produced it so two runs are comparable.

Verdicts

Verdict	Meaning
CLEAR	No real-world collision, or protected use.
CAUTION	Real entity exists; depiction is defensible but worth a look.
CHANGE_RECOMMENDED	Real entity + unflattering or illegal depiction.
LICENSE_REQUIRED	Rights holder identified; clearance must be purchased.
UNRESOLVED	Insufficient evidence. Surfaced honestly, never guessed.

UNRESOLVED is a first-class verdict. A clearance tool that claims certainty it does not have is worse than no tool.

---

Risk-routed search

Parallel pricing: fast at $1 / 1 000 requests, advanced at $5 / 1 000.

fast is enough to establish whether a neutrally-mentioned entity exists. advanced is spent only where the verdict is genuinely in play — entities shown unflatteringly, or high-exposure categories (songs, artworks, real people).

Five times cheaper across the bulk of the fan-out, with no loss of quality where it matters.

Measured: a full pass over a 100-page screenplay (~180 entities) costs about $0.20 and runs in minutes, against roughly a week and several thousand dollars for a manual clearance pass.

---

Quick start

py -3.12 -m venv .venv
./.venv/Scripts/python.exe -m pip install -r backend/requirements.txt
cp .env.example .env        # set PARALLEL_API_KEY and GOOGLE_CLOUD_PROJECT

Parse a screenplay:

PYTHONPATH=backend ./.venv/Scripts/python.exe -c "
from greenlight.ingest.fountain import parse_file
d = parse_file('samples/seventeen_minutes.fountain')
print(len(d.scenes), 'scenes')"

Offline fixture harness

Credits are finite. FIXTURE_MODE prevents paying twice for identical calls:

Mode	Behaviour
live	calls the API, stores nothing
record	calls the API and writes the response to disk
replay	reads disk only — no network, no credits

Record once, then develop entirely in replay. The test suite runs in replay, so CI is free and deterministic.

---

Test screenplay

samples/seventeen_minutes.fountain is a short screenplay written for this project and deliberately seeded with clearance landmines covering every report category — including one intentionally harmless trap the system must return as CLEAR.

That last one matters: it is the proof the system reasons about context instead of painting everything red.

Expected verdicts: samples/EXPECTED.md.

---

Architecture

Full network and cloud design: docs/ARCHITECTURE.md.

Browser (Material 3 SPA)
   │
   ▼
Cloud Run · gl-api ──── Cloud Tasks ────▶ gl-orchestrator (ADK agent runtime)
   │                                          │  fan-out, rate-limited
   │                                          ▼
   │                                     gl-research-worker ──▶ Parallel Search API
   ▼                                          │
Firestore ◀───────────────────────────────────┘
   ▲
   └── Vertex AI · Gemini (extraction + classification)

Built natively on the Agent Development Kit, as recommended by the hackathon resource guide, rather than a third-party wrapper framework.

---

Versioning & releases

This project follows Semantic Versioning and Keep a Changelog.

• Version of record: pyproject.toml → project.version
• History: CHANGELOG.md
• Releases are git tags vX.Y.Z

main is the production branch. Every push to main deploys. CI runs lint and the offline test suite on every push and pull request; a green run on main triggers the Cloud Run deployment.

---

Scope, stated honestly

GREENLIGHT does not replace the official clearance report required by an E&O insurer, and it is not legal advice.

It is upstream triage: catch problems during writing, when the fix is free, and hand the clearance vendor a script that is already clean.

---

License

Apache 2.0

_{Built for the Agentic Cinema hackathon · Parallel track}