How it works
What this is
First Pass is an operator console that checks a finished film master against a streaming destination’s delivery rules before you submit. Pure Python evaluates the master’s technical metadata; a single bounded agent then writes the operational follow-through into Grafana — incidents, annotations, and alert rules — when blockers appear.
The problem
Industry commentary from post-production delivery specialists puts first-submission QC failure in the 20–30% range — a directional figure, not an official published statistic. Most often the cause is mundane and preventable: an audio mix delivered at theatrical loudness (~−24 LUFS) against a streaming platform spec requiring ~−27 LUFS. Every rejection incurs redelivery fees and risks missing an announced premiere date.
How it works
A deterministic check engine compares the master to the selected destination spec and returns structured findings. Python publishes the Delivery Readiness dashboard and telemetry directly. The agent never computes those measurements — it only acts through four allowlisted Grafana tools when the engine reports blockers.
How to read the console
Verdict banner. After RUN finishes, the burn-in frame shows PASS (green) or REJECT — N blockers (red). The destination, languages, and run date sit in the corners; elapsed timecode ticks at 24 fps while a run is active.
Loudness meters. Bars are relative, not absolute. 0 LU is this destination’s loudness target (for StreamOne that is −27 LUFS; for HallArc, −24 LUFS). A bar left of centre is quieter than required; right of centre is louder. True Peak sits on a separate shared axis against the destination’s ceiling.
Fix plan. Each row is a non-conformance: clause, severity, measured vs expected, language, and finding. Blockers stop delivery; warnings do not. Jobs are ordered by how long the fix takes to arrange — longest lead time first — not by processing order.
Language readiness. Per-language scores show how close each version is to shippable. For a simultaneous multi-language release, certification of the original language can gate clearance of the dubs, so one pending certificate can hold up every other version.
Action ledger. Every agent write to Grafana appears here with time, operation, detail, and a link into the stack. On a PASS the agent writes nothing.
The two destination specs
StreamOne and HallArc are two fictional destinations authored for this project, with deliberately different targets — which is why the same master can pass one and fail the other.
- StreamOne — streaming-style loudness at −27 LUFS (±2 LU), True Peak ceiling −2.0 dBTP, wide-gamut HDR (BT.2020).
- HallArc — theatrical-style loudness at −24 LUFS (±1 LU), True Peak ceiling −1.0 dBTP, BT.709 picture.
A theatrical mix near −24 LUFS can clear HallArc and still fail StreamOne on loudness alone.
Guardrails
- One bounded Google ADK agent — not a multi-agent crew.
- Four allowlisted MCP tools only: create_incident, add_activity_to_incident, create_annotation, alerting_manage_rules.
- Deterministic Python evaluates every clause; the model never computes measurements.
- The Delivery Readiness dashboard is published by Python over REST (POST /api/dashboards/db), not by the agent.
Links out