Data & Privacy

1. Your work, your IP

• Filmmaker uploads, scripts, screeners, and trailers remain your intellectual property.
• Your work is never used to train generative AI models — Anthropic's commercial API terms apply to every analysis call PACCS makes.
• Your work is never resold, redistributed, or shared outside the festival programming workflow you explicitly authorized.

2. What we store and why

• Metadata (title, synopsis, runtime, country, genre, etc.) — required for analysis and matching.
• Transcripts — generated from your trailer audio so the agents have richer context than synopsis text alone. Cached on your film row; never re-transcribed for the same row.
• Cinematography, audio, and engagement findings — generated from your video for craft analysis. Stored alongside your film row.
• Decision events — programmer decisions paired with PACCS scores at the moment of decision. We use these to measure how accurately PACCS predicts each festival's actual taste.
• Calibration aggregates — anonymized statistics derived from decision events that sharpen recommendations over time. Aggregate-only; no individual film identifiers leak between festivals.

3. What we don't do

• No persistent storage of your screener. PACCS streams from the hosting link you provided (YouTube, Vimeo, FilmFreeway, etc.) — we do not download or retain your video file.
• No resale of data, ever. Not to advertisers, not to data brokers, not to any third party.
• No model training on filmmaker work. PACCS uses commercial Anthropic / OpenAI APIs whose terms prohibit training-from-input.
• No automated platform scraping. The FilmFreeway URL importer respects robots.txt and identifies itself with an honest User-Agent. The festival CSV import is operator-driven, never background.

4. How PACCS learns

PACCS does not get smarter through opaque model training on your films. It gets smarter through a transparent measurement loop:

programmer decision → measure prediction accuracy → improve recommendations

The historical 4,726-film corpus seeded the agents. Every subsequent programmer decision becomes a labelled outcome that lets us report — and defend — how accurately PACCS predicts each festival's taste. The math is Pearson correlation on (PACCS Score, decision) pairs; the result is public on each festival's dashboard at n ≥ 20 decisions.

This is calibration, not training. The agents themselves are not retrained on filmmaker work.

5. PACCS Open vs PACCS Secure

Every analysis runs through the same multi-agent pipeline. The distinction is in the source URL and the audit trail required to access it.

6. Your controls

• Opt out at submission time via the festival's AI-Assisted Programmer Review custom question on their FilmFreeway submission form. PACCS respects each filmmaker's choice on a per-row basis.
• Request data deletion by emailing peekaboon@pahus.org. We respond within 14 days.
• Withdraw consent at any time and have your analysis records anonymized. Decision aggregates that include your film stay in the corpus only in non-identifiable form.

7. Compliance

• UK company. PAHUS Studios Limited (Co. 13592044) and PAHUS CIC (Co. 13158127), UK-headquartered.
• Postgres data is hosted in Oregon, US (Render). EU-based customers may request alternative hosting once volume justifies provisioning a separate region — write to peekaboon@pahus.org.
• GDPR-aligned data subject rights honored. Access, rectification, erasure, portability, and objection requests are all supported via the contact below.