How does testing an AI director agent on an unfamiliar genre prove the style was truly learned?

Run the agent on a genre the director never shot. Because the content and genre tropes can't carry the output, anything stylistic that survives — shot length, lensing, palette logic, blocking grammar, sound architecture — has to come from what the agent actually learned. If it asks the right clarifying questions and still produces coherent style, the bible is grammar, not mimicry.

Pick a genre clearly outside the director's filmography — a courtroom thriller for a horror director, a heist for a romance auteur — and brief the invideo agent at a single sentence of premise. The invideo agent is an agentic video creation tool that holds a loaded director bible across every shot, so this test reads how deeply that bible got internalized.

What 'pass' looks like. The agent asks clarifying questions before generating — era, threat, location, deliverable — rather than auto-filling with horror-coded defaults. In one documented test, the agent was asked for a courtroom thriller through a James Wan lens (a genre Wan has never made); it asked about era and the nature of the threat first, then pulled a named principle from page 12 of the bible — "Mood Over Narrative, the substitution rule" — and applied it to a scene type the document had never addressed. That is grammar transfer, not pattern-matching.

The low-level markers that have to survive. Cross-genre authorship attribution research has long shown that what distinguishes a director's style from content lives in low-level formal markers — shot length distribution, lens choice, lighting ratio, palette, framing angle. Check those, specifically. For the Wan test the bible's 85:15 dark-to-light ratio, 2.40:1 hard-matte framing, and spherical lensing (no anamorphic flare, circular bokeh) should appear in a courtroom interior with no horror cues. For a Wong Kar-wai cross-genre test, the named tonal modes with exact hex values and the doorway-static-hold closing structure should surface autonomously. If the markers vanish the moment the genre changes, the agent learned topics, not style.

Stress-test checklist to run. Give the agent four things and watch: (1) a one-line premise in an off-genre — no references, no mood board; (2) a request for a shot list, so you can audit shot-length rhythm and lens grammar against the bible; (3) a single frame generation, so you can audit palette, lighting ratio, and composition; (4) a sound/architecture brief, since half of what makes some directors' work land is what the audience hears before what they see — a real bible encodes that and the agent should apply it unprompted. Score each output against the bible's named rules; deviations the agent flags itself ("shadows are leaning blue-green instead of neutral gray, pulling Stage A rule") are the strongest signal that the rules are live in context.

Why this works as a validation step. In-genre tests are contaminated: a horror prompt will produce horror-coded output even from a thin bible, because the genre tropes do the work. Strip the genre away and only the learned formal language remains to explain the result. Use the off-genre test as the final gate before committing budget — across documented productions, finished AI shorts ran $750–$5,000 and 2–5 days, so catching a shallow bible before that spend is the point of the test.

Watch some of these to see what works for you:

Before generating a single frame, I stress-tested the doc. I asked for a courtroom thriller through the James Wan lens. Something he's never made. If the agent was just mirroring style superficially, it would fail here.

— invideo's creative team