CUTLINE turns live production evidence into a bounded recovery proposal, keeps the decision with a human, executes only through an authenticated Google Cloud boundary, and proves the outcome with fresh telemetry.
Launch CUTLINE · Watch the 2-minute demo · Read the submission package
Agentic cinema can create shots faster, but a production still has to answer a hard operational question:
Will the final shots cross package lock—and can the team prove a safe recovery when they will not?
CUTLINE is not an AI movie generator. It is a release-assurance control plane for the last mile of agentic production: evidence, diagnosis, human authority, bounded action, and verification.
The included Eclipse Protocol / SQ-42 scenario is a controlled, deterministic demonstration with 18 final VFX shots, 4,800 frames of backlog, and 24 minutes until package lock. The workload is synthetic; the hosted Grafana MCP evidence path, Gemini synthesis, and Google Cloud action boundary are real and disclosed.
| Stage | What CUTLINE does | Authority boundary |
|---|---|---|
| 1. Investigate | Queries Prometheus and Loki through the official Grafana MCP and binds every item to the active run. | Provider timestamps and run labels are required; stale or mismatched evidence is rejected. |
| 2. Synthesize | Gives Gemini 2.5 Flash a bounded evidence packet with observations, alternative, discriminator, and falsifier. | The model cannot approve, execute, own authoritative arithmetic, or declare success. |
| 3. Decide | Shows deadline math and one narrow, reversible recovery proposal. | A named human operator must explicitly approve or reject it. |
| 4. Execute | Sends the approved allowlisted plan through an authenticated Cloud Run action boundary. | Idempotency and Firestore atomic-create semantics prevent duplicate mutation. |
| 5. Verify | Retrieves later Grafana evidence and evaluates six deterministic recovery gates. | Execution is not success; only fresh evidence can produce VERIFIED. |
flowchart LR
G["Grafana MCP<br/>Prometheus + Loki"] --> E["Evidence packet<br/>run-bound + time-bound"]
E --> M["Gemini 2.5 Flash<br/>bounded synthesis"]
E --> D["Deterministic<br/>deadline math"]
M --> P["Recovery proposal"]
D --> P
P --> H{"Human approval"}
H -->|Approved| A["Authenticated<br/>Cloud Run action"]
H -->|Rejected| X["No mutation"]
A --> V["Fresh Grafana<br/>verification"]
V --> R["Receipt + audit lineage"]
| Signal | Before recovery | After recovery |
|---|---|---|
| Render throughput | 120 frames/min | 320 frames/min |
| Projected completion | 16 minutes late | 9 minutes early |
| CUDA OOM state | Active | Cleared |
| Verification gates | Not evaluated | 6 / 6 passed |
| Final state | Action required | VERIFIED |
The result is evidence from a controlled scenario—not a benchmark claim about an external studio pipeline.
View the verified command center
| Layer | Technology | Responsibility |
|---|---|---|
| Agent runtime | Google ADK + Gemini 2.5 Flash on Vertex AI | Evidence-grounded, bounded synthesis |
| Workflow service | FastAPI + Pydantic | State machine, impact arithmetic, approvals, receipts, and verification |
| Observability | Official grafana/mcp-grafana |
Live Prometheus and Loki queries through discovered MCP schemas |
| Action boundary | Google Cloud Run | Authenticated, allowlisted, idempotent recovery execution |
| Persistence | Firestore | Durable atomic action idempotency |
| Secrets | Google Secret Manager | Runtime-only provider and action credentials |
| Interface | Accessible vanilla web UI | Sequence-first release command and complete evidence disclosure |
- The browser receives no Grafana, Gemini, action, or Google Cloud credentials.
- The private Grafana MCP service uses Google service identity and read-only tools.
- A separate write-only credential publishes bounded scenario telemetry.
- The model can synthesize evidence but cannot mutate state or certify success.
- Every approval, action, evidence item, and verification result is bound to one run ID.
- Missing providers, identities, datasources, or action configuration fail closed.
- Open the live Cloud Run app.
- Select Investigate with Grafana evidence.
- Review the evidence-backed diagnosis, alternative, falsifier, and deadline math.
- Approve the bounded recovery as the named operator.
- Execute the approved plan.
- Verify with fresh evidence and inspect the receipt and audit lineage.
No private GCP or Grafana credentials are required for the judge-facing flow.
Prerequisites: Python 3.12+, uv, and
agents-cli.
git clone https://github.com/aitrailblazer/cutline.git
cd cutline
uv sync --dev --extra lint
uv run uvicorn app.fast_api_app:app --host 127.0.0.1 --port 8011Open http://127.0.0.1:8011. Local mode uses deterministic adapters, is clearly
labeled LOCAL CONTROLLED, and makes no live-provider claim.
Copy .env.example to .env and provide deployment values through Secret
Manager. Never commit secrets.
CUTLINE_MODE=live
GRAFANA_MCP_URL=https://<private-mcp-cloud-run-service>
GRAFANA_MCP_TOKEN=
GRAFANA_MCP_USE_GOOGLE_ID_TOKEN=true
GRAFANA_MCP_AUDIENCE=https://<private-mcp-cloud-run-service>
CUTLINE_ACTION_URL=https://<cloud-run-action-service>
CUTLINE_ACTION_TOKEN=<secret-manager-injected-action-token>
The Grafana endpoint must expose the official list_datasources,
query_prometheus, and query_loki_logs operations. CUTLINE discovers explicit
datasource UIDs and rejects missing, stale, malformed, or run-mismatched
results.
deploy/grafana-mcp/Dockerfile pins the official Grafana MCP image used by the
private Cloud Run evidence boundary. Write tools are disabled in that service.
./scripts/test
agents-cli lint
node --check app/web/app.js- 51 automated tests
- 732 deterministic statements at 100% coverage
- 128 deterministic branches at 100% coverage
- 43/43 user-visible feature stories verified
- 100 Lighthouse scores for accessibility, best practices, SEO, and agentic browsing in the final hosted audit
The canonical audit source is
feature_status_tracker.csv. Its generated views
are feature_status_tracker.xlsx and
feature_status_tracker.html. The final closure
gate is preserved in
qa_evidence/commands/competition_audit_final_closure_gate_2026_08_02.txt.
app/ Workflow, adapters, agent runtime, API, and web UI
deploy/grafana-mcp/ Pinned official Grafana MCP Cloud Run image
docs/ Product spec, execution plan, and submission package
qa_evidence/ Reproducible test, browser, cloud, and media evidence
tests/ Unit, integration, workflow, and evaluation coverage
feature_status_tracker.* Canonical audit plus generated human views
- Winning product specification
- Execution plan
- Competition submission package
- Feature QA report
- Changelog
The public hosted revision runs with Gemini 2.5 Flash on Vertex AI, a private official Grafana MCP Cloud Run service, and an authenticated Google Cloud action boundary.
gcloud config set project <dedicated-cutline-project>
agents-cli deployLocal tests never deploy infrastructure, mutate billing, or start a managed paid evaluation.
Licensed under the Apache License 2.0.
Evidence. Human authority. Bounded action. Proof.