Provenance Index Methodology
A plain-language standard for measuring generative AI and live-action filmmaking across vertical streaming platforms. Written for journalists, creators, platforms, and industry analysts.
1. The Three Buckets (Never Two)
Most binary AI detection indices make a fatal methodological error: they assume that anything not proven to be AI must be human, or vice-versa. In an industry releasing tens of thousands of micro-episodes monthly across dozens of mobile apps, incomplete data is a reality.
Vertical Dramas Official strictly partitions every title into one of three independent buckets:
Verified human cast, physical location cinematography, human voice acting, and original writing with confidence ≥ 0.80.
Synthetic voice cloning, LLM-generated screenplays, neural face swaps, or fully synthesized video generation with confidence ≥ 0.80.
Titles undergoing primary source verification, disputed entries, or titles with detection confidence below 0.80.
coverage = (confirmed_live_action + ai_involved) / total_titles
The Cardinal Rule: We never fold unconfirmed titles into either side. Doing so would artificially inflate or deflate platform percentages and transform a rigorous metric into speculative editorializing.
2. Signals Moving a Title to AI-Involved
To assign an AI-involved verdict, our classification pipeline requires verifiable signals corroborated by production filings, audio spectral analysis, casting agency registries, or verified platform disclosures.
- Synthetic Voice Dubbing (Confidence ≥ 0.85): Neural speech synthesis replacing original actor voices, identified via harmonic unnaturalness, cadence repetition, or platform voice provider attribution.
- Generative Screenplay (Confidence ≥ 0.80): Scripts explicitly registered as AI-assisted or machine-generated in copyright filings, studio trade press announcements, or production credits.
- Neural Face Replacement (Confidence ≥ 0.90): Digital face swapping onto body doubles, verified via frame-level temporal boundary inconsistencies or performer opt-out declarations.
- Full Video Generation (Confidence ≥ 0.95): Works generated wholly via text-to-video diffusion or autoregressive video foundation models (e.g. Sora, Kling, Runway, Hailuo).
3. The Statistical Suppression Rule
Publishing headline percentages on sparse sample sizes is reckless. If an app has only 8 titles in our catalog and 7 happen to be live-action, publishing “88% Live Action” would present a false picture of the platform's true operational footprint.
Therefore, we enforce a strict Publication Threshold:
- At least 50 titles have been thoroughly classified with confidence ≥ 0.80.
- At least 40% of the platform's known catalogue has been classified.
When a platform does not meet these criteria, we do not guess, estimate, or assign letter grades. We display:
“Not enough data yet — {n} of {total} titles classified”
4. Nightly Recomputation & Right of Reply
The index is recomputed automatically every night at 00:00 UTC. As our researchers verify new titles and catalog imports expand, platforms automatically graduate from suppressed to published status when their data meets the threshold.
Right of Reply & Correction Policy:
Streaming platforms, production studios, directors, and cast members have an absolute right of reply. If any title is misclassified or if production telemetry has been misinterpreted:
When a dispute is submitted with supporting evidence, an inline badge (“Under dispute”) is immediately stamped on the title across all catalog pages until independent moderation review is complete.