What R00 was
One scene — the elevator, the woman in the red jacket — generated by a handful of configurations, reduced to eight blind samples, reviewed by one person under a seven-rating instrument and then judged by three AI models, audiovisually and from the audio alone.
Finding 1 — seven ratings were too many
The seven 1–5 ratings, several overlapping, felt longer and more redundant than they should. Nothing was re-scored: the eight R00 records stay exactly as collected. But the finding drove Core R01's shape — three factual questions and three plain-language ratings instead of seven dimensions — as a new, versioned instrument.
Finding 2 — a clip's file could name its maker
After the human pass was locked, the AI judging tool refused to send five of the eight video files: their containers carried content credentials, a watermark assertion and, in one, workflow metadata naming a model repository. Nothing in the locked review appears to have been affected, but the gap was real. Core R01 strips every metadata box from every clip by one uniform transformation, pads every file to one size, and scans what it sends.
Finding 3 — an AI contract that contradicted itself
The second AI judging run completed all 48 requests cleanly and still produced only 9 valid records. The contract printed a blank confidence value while the schema required an integer, so the judges improvised their own scales. The fix was to constrain the response at the provider with a schema generated from the frozen record schema. Core R01 goes further: the judges' structured output enforces the same not-applicable rules the human questionnaire does.
Finding 4 — agreement is not accuracy
With the contract fixed, the AI panel was useful for detecting whether the event happened and systematically unreliable about who produced it. Across the rated dimensions it ran almost a full point above the human on average; the audio-only judges reported high confidence on every record while disagreeing widely with each other. A lenient judge can still track a human ranking. So in Core R01, AI judges answer the same factual questions as people, and the comparison that counts is factual agreement — especially on source and cause — not rating correlation.