AIVideoRunbook: The Logbook That Turns Three Quality Gates Into One Pipeline

You can have three great gates and still lose track. The madebysaira safety net already existed as three separate CLIs:

• AIVideoCreditGuard checks the job before you pay — lints the prompt, estimates real cost, flags consistency and morph risk.
• AIVideoQualityGate checks the file after render — ffprobe for resolution, codec, fps, duration, audio.
• AIVideoAdherenceGate checks what the clip actually shows — motion health, morph-drift, lip-sync, creative-contract adherence.

Run them in a shell loop, sure. But then you have: which clip passed which check? which model produced it? which generation was the one that almost worked? And when a client asks "where's the version with the blue scarf," you dig through folders named after timestamps and hope.

That is the gap Runbook fills.

AIVideoRunbook

A small, offline-first CLI that turns the three gates into a single, ordered pipeline — and remembers everything.

runbook run clip.mp4 --prompt-from prompt.json → exit 0, 1, or 2.

CreditGuard  →  Runbook (orchestrate + log)  →  QualityGate  →  AdherenceGate  →  VERDICT + CATALOG

The command runs all three checks in sequence, then writes the result — provider, model, seed, prompt, cost, per-gate verdicts, and exit code — to a local SQLite catalog (runs.db). Exit 0 = deliver, 1 = eyeball, 2 = regenerate. It mirrors the exit-code contract your gates already use, so it drops straight into any IF-node.

One command, three jobs

Where the gates each answer one question, Runbook answers: "did the whole job work, and can I prove it?"

The catalog stores the full chain: who paid (CreditGuard), what the file said (QualityGate), and what it actually showed (AdherenceGate). No cloud, no API keys, no accounts — just a SQLite file you can cp or rsync like any other asset.

Why not just shell scripts?

You can chain three if statements. Runbook does a few things the one-liners don't:

• The verdict is authoritative — it walks the gates in the right order and the worst severity wins. A warn on QualityGate followed by ok on AdherenceGate is still warn-overall, not silently ok.
• The log is queryable — every clip gets an @id, a timestamp, and a --json record. No more "which generation used that prompt with the Veo hands fix?"
• It is n8n-native — the --json flag + exit codes let you wire the whole pipeline into a workflow. 0 → deliver to client, 1 → ping me, 2 → regenerate. I included a ready-to-import examples/runbook.n8n.json.
• It degrades gracefully — if a gate binary isn't installed, Runbook skips it and keeps going. You get the catalog benefit immediately; install gates as you adopt them.

The three-gate stack, now connected

This is the fourth pillar, not a replacement. CreditGuard lints before you burn the credit, QualityGate checks the bytes, AdherenceGate checks the pixels — and Runbook is the logbook that remembers what happened at each step so you never have to.

The full stack now reads:

Before you pay → CreditGuard
After you render → QualityGate (technical) + AdherenceGate (semantic)
Every time → Runbook catalogs the verdict

Try it

• Repo: github.com/madebysaira/AIVideoRunbook (Python 3 + ffmpeg, zero required pip deps, 7 tests passing)
• Install: pip install git+https://github.com/madebysaira/AIVideoRunbook.git
• Quick start:

  runbook new --provider Kling --model kling-vid-1.0 \
    --prompt "dolly zoom hallway, slow push" --seed 42 --budget 1.50
  runbook run render.mp4 --prompt-from prompt.json --json
  runbook find "dolly"
  

• n8n: import examples/runbook.n8n.json → your gates, wired to a catalog.

Clone it, run your next clip through, and search for it by lunchtime. Three gates were never the whole safety net — the logbook is what makes the net close.", "repo": "https://github.com/madebysaira/AIVideoRunbook"