OpenPandora is a small production operation — a YouTube network and restaurant software, run day to day by a team of AI assistants ("the Court") under one accountable human owner, Manuel Avila. We are not a research lab; we make shows and tools for a living. What follows are field notes from that work: two write-ups of systems we built because production broke without them, what they caught, and — just as carefully — what they did not prove. Everything below was produced by the AI team named in each byline, measured in production, and is published with its limits stated in the text. Neither paper is peer reviewed, and neither claims more than its receipts support.
The Ear Equation: Machine-Checkable Audio Gates for an AI-Generated Video Pipeline, Calibrated Against One Producer's Ear
Our episodes are generated clip by clip and assembled by software, and audio defects — dead-air holes, phantom words pasted into silence, hard joins, a missing music bed — became the dominant quality failure: masters that measured "correct" on loudness standards still failed when the producer listened. This paper describes the Ear Equation, a set of machine-checkable gates on joins, pauses, music-bed presence and loudness, drawn from the defects the producer actually heard and enforced by scripts that refuse a master that has not passed. It reports one real failure (a speech-only master at −14.12 LUFS, rejected by ear: no bed, 77 hard joins, a −57.7 dBFS pause floor) and its clean rebuild from first-generation clips, which passed every machine gate (−14.01 LUFS, −1.81 dBTP). It also reports the limits plainly: at the study cutoff the calibration ledger had three rows, and no producer PASS has yet been paired with a machine PASS. An operational case study — not a perceptual study, and no model was trained.
Byline: Manuel Avila (Founder & Owner) and the Court — OpenPandora, San Marcos, CA. Manuel Avila is the accountable human author; the AI seats are credited contributors.
One producer runs a media operation through thirteen AI assistant seats spread across several products and computers, coordinated through one shared file shelf. Session resets lose working context mid-task, and a seat's report of its own work is often wrong — so the operation runs on files, not on trust: per-seat boot cards read first after any reset, an end-of-turn state footer, append-only skill locks tied to enforcing scripts, and file-evidence acceptance, under which nothing counts as done until it is a file on the shelf with a byte count and a SHA-256. The paper logs what those mechanisms caught in five days of production (Oct 26–30, 2026) — including a Short nearly cut from the wrong show's master and an 18.25 GB only-copy master queued for deletion — and where they are still weak. It is an incident log from one operation, not a controlled experiment, and it claims no causal effect.
Byline: Manuel Avila (Founder & Owner) and the Court — OpenPandora, San Marcos, CA. Manuel Avila is the accountable human author; the AI seats are credited contributors.