AI structures screenplay changes. Deterministic safety rules decide whether the camera can roll.
A proof-first safety review cockpit for film productions.
Live app | Demo video | Architecture | Submission copy
Google ADK
Gemini
FastAPI
Python safety engine
Grafana MCP
Vercel
A late screenplay revision can introduce a firearm, smoke, a confined space, a stunt, or powered equipment after the production plan is already in motion. Universal CallSheet turns that revision into a reviewable production record:
changed text -> affected departments -> hazard evidence -> stage decision -> required clearances -> receipt
The memorable outcome is intentionally simple: when a critical hazard is present, the system says STOP, explains why, and shows what must happen next.
Use the header links in this order: demo video for the plain-English story, live app for the S2 refusal path, then architecture notes for the source-level explanation.
For judges: watch the demo first, then use the live app to verify the same refusal outcome.
flowchart LR
A["S2 screenplay revision<br/>firearm + smoke + confined space"] --> B["UCS structures<br/>what changed"]
B --> C["Python safety engine<br/>checks governed hazards"]
C -->|critical hazard| D["STOP<br/>set frozen"]
D --> E["Required clearances<br/>before rehearsal or roll"]
D --> F["Receipt<br/>JSON + Grafana state"]
classDef input fill:#17212b,stroke:#8ed5ff,color:#ffffff,stroke-width:2px;
classDef process fill:#202a33,stroke:#8ed5ff,color:#ffffff,stroke-width:2px;
classDef stop fill:#8f2d35,stroke:#ff7b74,color:#ffffff,stroke-width:3px;
classDef proof fill:#2b2419,stroke:#ffb95f,color:#ffffff,stroke-width:2px;
class A input;
class B,C process;
class D stop;
class E,F proof;
This is the product in one sentence: AI may interpret the revision, but deterministic permission is required before the camera can roll.
Use the live app from the header and select S2 in the scenario deck. The proof surface is the same in the demo and in the deployed product: a concrete revision, a deterministic STOP, required clearances, and an inspectable receipt.
- Open the live app and keep
Auto-reviewenabled. - Select
S2: Confined Space Prop Firearm Shootout. - Watch the revision move through structured analysis and the deterministic safety engine.
- Read the
STOPdecision, affected teams, and required clearances. - Open the evidence panels and raw JSON to verify the result instead of trusting a claim.
| What you will see | Why it matters |
|---|---|
| S2 adds a blank firearm discharge and atmospheric smoke | A concrete production change, not an abstract AI demo |
DiffOutput, CascadeOutput, and HazardTagOutput |
The model work is structured and inspectable |
STOP from the Python safety engine |
Final authority is deterministic and testable |
| Required armorer, safety, and rescue clearances | The result creates an operational next step |
| Raw JSON and receipt state | The visible story maps back to a backend response |
flowchart LR
A["Screenplay revision"] --> B["ADK / Gemini<br/>structured extraction"]
B --> C["Pydantic<br/>contract validation"]
C --> D["Deterministic<br/>Python safety engine"]
D -->|GREEN| E["Standard checks<br/>may continue"]
D -->|RED / STOP| F["Hold camera<br/>require clearances"]
D --> G["Grafana MCP<br/>receipt when configured"]
F --> H["Frontend proof HUD<br/>JSON, evidence, next action"]
G --> H
classDef input fill:#17212b,stroke:#8ed5ff,color:#ffffff,stroke-width:2px;
classDef ai fill:#202a33,stroke:#8ed5ff,color:#ffffff,stroke-width:2px;
classDef gate fill:#1d3027,stroke:#56e5a9,color:#ffffff,stroke-width:3px;
classDef hold fill:#8f2d35,stroke:#ff7b74,color:#ffffff,stroke-width:3px;
classDef proof fill:#2b2419,stroke:#ffb95f,color:#ffffff,stroke-width:2px;
class A input;
class B,C ai;
class D gate;
class E,F hold;
class G,H proof;
The model is useful for interpreting ambiguous screenplay language. It is not trusted with the final permission decision.
| Layer | Responsibility | Evidence in this repo |
|---|---|---|
| Google ADK / Gemini | Extract changes, department impact, and governed hazard tags | src/agent/pipeline.py, src/agent/models.py |
| Pydantic contracts | Reject malformed structured output | src/agent/models.py, API tests |
| Python safety engine | Map hazard rows to GREEN, REVIEW, RED, or STOP |
src/engine/safety.py |
| Grafana MCP adapter | Publish significant results as an annotation when configured | src/engine/grafana_client.py |
| Frontend HUD | Show the decision, evidence, provenance, and next action | public/index.html, public/app.js |
The frontend checklist records sign-off progress; it cannot override a backend STOP verdict. That rule is covered by a browser regression test.
flowchart TD
A["POST /api/analyze<br/>src/main.py"] --> B["RevisionPipeline<br/>src/agent/pipeline.py"]
B --> C["1. DiffOutput<br/>changed screenplay text"]
C --> D["2. CascadeOutput<br/>affected departments"]
D --> E["3. HazardTagOutput<br/>governed hazard rows"]
E --> F["evaluate_safety<br/>src/engine/safety.py"]
F -->|GREEN| G["Proceed with standard checks"]
F -->|REVIEW / RED / STOP| H["Hold camera<br/>required clearances"]
F --> I["publish_to_grafana<br/>src/engine/grafana_client.py"]
G --> J["final_output JSON"]
H --> J
I --> J
J --> K["renderDashboard<br/>public/app.js"]
classDef api fill:#17212b,stroke:#8ed5ff,color:#ffffff,stroke-width:2px;
classDef analysis fill:#202a33,stroke:#8ed5ff,color:#ffffff,stroke-width:2px;
classDef gate fill:#1d3027,stroke:#56e5a9,color:#ffffff,stroke-width:3px;
classDef hold fill:#8f2d35,stroke:#ff7b74,color:#ffffff,stroke-width:3px;
classDef output fill:#2b2419,stroke:#ffb95f,color:#ffffff,stroke-width:2px;
class A api;
class B,C,D,E analysis;
class F gate;
class G,H hold;
class I,J,K output;
| Stage | Runtime responsibility | Source |
|---|---|---|
| 01 | Resolve a preloaded or custom scene and accept the revision request | src/main.py |
| 02 | Produce schema-constrained diff, cascade, and hazard outputs | src/agent/pipeline.py |
| 03 | Apply hazard-row rules without an LLM in the decision loop | src/engine/safety.py |
| 04 | Attempt the official Grafana MCP annotation for significant outcomes | src/engine/grafana_client.py |
| 05 | Return one JSON record and render the proof HUD | src/main.py, public/app.js |
The same path is available in live Gemini mode and in the clearly labeled local fallback. Fallback keeps the demo reviewable; it does not pretend to be a Gemini receipt.
The local Grafana path was verified with a Backlot Safety Wall showing scene severity and open locks. Hosted Grafana publishing is supported, but the live app labels it as published only when the current backend request returns a receipt.
| Capability | Status | Source of truth |
|---|---|---|
| Vercel deployment | Verified | Live app link in the header |
| Public demo video | Verified | Demo video link in the header |
| S2 refusal path | Tested | tests/ui-smoke.spec.js |
| Deterministic safety decision | Tested | tests/test_safety_engine.py, src/engine/safety.py |
| Structured analysis path | Tested and fallback-safe | tests/test_agent_pipeline.py, src/agent/pipeline.py |
| Auto-review workflow | Tested | tests/ui-smoke.spec.js, src/main.py |
| Local Grafana MCP write path | Verified locally | src/engine/grafana_client.py, docs/wall.png |
| Hosted Grafana MCP | Supported, unverified | Requires a configured hosted MCP endpoint and returned receipt |
| Public repository | Pending owner action | The repository owner must change GitHub visibility to public |
The app distinguishes live requests, structured fallback, history snapshots, static snapshots, and Grafana receipt state. No live Gemini or hosted Grafana claim should be inferred when the corresponding status is not visible in the current response.
Current verification: 21 backend tests and 8 responsive browser tests pass on the submission branch.
- This is an assistive hackathon prototype, not legal advice and not a replacement for qualified safety leadership or regulators.
- The governed hazard mapping currently targets the Alberta OHS model encoded in
src/engine/safety.pyand its supporting data files. - SQLite history on Vercel is best-effort and ephemeral; the live POST response is the source of truth.
- Hosted Grafana publishing is not claimed until a configured endpoint returns a verified MCP receipt.
This submission is maintained from a collaborator lane covering the current ADK/Gemini agent and product implementation, deterministic safety integration, Grafana MCP path, reliability work, deployment, tests, and judge-facing materials. The repository owner provides the upstream production-safety foundation, domain and jurisdiction data, repository governance, final merge control, credentials, and GitHub visibility. The repository history supports one shared hackathon solution: AI structures the revision, deterministic rules decide, and the crew receives a proof-backed next action.
- Watch the demo video
- Launch the live app
- Read the architecture
- Open the submission copy
- Explore the GitHub repository
Submission readiness: the hosted app, public demo video, code-flow documentation, test evidence, and local Grafana proof are prepared. The repository owner must make the GitHub repository public before final submission.