Integrity: make it trustworthy

Professional audiences forgive plain design. They don't forgive a wrong number. The most common criticism of 2026's viral AI-made videos was that nobody fact-checked them. ClearFrame treats verification as part of the render.

This is a general professional standard, not legal advice. If your organisation works under specific rules for public communications, route the storyboard and final render through that review. The storyboard plus its sources list is designed to be the audit trail.

1. Every figure has a source

• Record each one in storyboard.sources: { "claim": "Revenue grew 42%", "source": "FY2026 annual report, p.12", "asOf": "2026-06-30" }.
• Keep figures in block props, never typed from memory (see clearframe-dataviz). Numeric blocks refuse to render without a visible props.source and a sources entry.
• Put a source line on screen for every data scene (props.source), with the period or as-of date.
• Illustrative numbers are labelled once, as what they are ("Example: a $60,000 salary", "Hypothetical illustration"), in the source line and the storyboard.
• Attribution is short and in the film's own words. Say where a figure comes from ("Alder Street Retirement Plan, 2027", "Sam Ortiz on Money at Work"), never what it is not. A rule of thumb said on air is attributed to the speaker, and that is its label. Do not write "not research", "illustrative only", "said on air; not advice" or "bars drawn to that rate" on screen: an honest scale needs no annotation, and a disclaimer on every scene reads as fear, not care.
• Round honestly. Precision should match the source, and "about" in the narration is fine when the screen shows the exact figure.
• Check narration against the screen: the number spoken, the number shown and the number in sources must agree.

2. Claims are balanced

• If you state a benefit, state its material limitation or risk with similar prominence (size, duration, position). Don't leave it only in a fine-print footnote.
• Use no guarantees or promissory language about uncertain outcomes. Use calibrated verbs (likely, about, tends to).
• Use no superlatives ("best", "fastest", "#1") without a named, dated basis on screen.
• Comparisons name what's compared: the same period, the same basis, the same units.
• Past results are not presented as predictions.

3. Disclosures people can actually read

• Source lines render at 22 px; a disclosure viewers must actually read belongs in a callout or statement (32 px and up) in the ink color, with contrast ≥ 4.5:1 (check warns on low-contrast palettes).
• Hold time is at least words ÷ 3 + 1 seconds, and never under 2.5 s. Set the beat's duration or hold to at least that.
• If a disclosure is required, say it in the narration too, not only on screen.

4. Accessibility

• Captions are burned in for every narrated film by default, to a broadcast standard (two lines of at most 42 characters, 28 on a vertical frame, timed to the measured words; docs/speech.md). Keep them on; captions: false is for a deliverable whose player adds its own. Export the same cues as SRT/VTT for every deliverable (clearframe captions). check warns when a caption covers the picture's type.
• Never encode meaning in colour alone. Pair colour with a sign, an arrow or a label.
• No flashing: nothing faster than 3 flashes per second, and no full-frame strobes.
• Narration should describe the key visual ("the line doubles by year three"), so the film works for listeners too.

5. AI transparency

• Voice: Gemini TTS is synthetic. Say so in the delivery notes and, where the audience would care, in the end card or description ("Narration: AI-generated voice").
• Imagery and footage from Gemini image, Lyria and Veo carry SynthID watermarks. Don't strip or obscure them, and label generated imagery when it could be mistaken for real documentation.
• Never generate or imitate real people (faces or voices), real logos, real documents, receipts, screenshots or data "evidence". Never clone a real person's voice without explicit consent.
• Generated visuals are illustration, never evidence.

6. Rights

• Fonts ship with the engine under open licences (Inter, Instrument Serif, IBM Plex Mono, Architects Daughter: SIL OFL 1.1, pinned google/fonts revisions with hashes in film/assets/fonts/provenance.json).
• Don't prompt music "in the style of" a named artist. Lyria blocks it, and you shouldn't want it.
• Only use brand assets (logos, marks, colours) the requester owns or is authorised to use.

7. Pre-delivery review (do all of it)

1. clearframe check <dir> shows 0 errors, and the missing-sources warning is resolved.
2. Read the vo lines in order and underline every factual claim. Each one appears in sources.
3. Read the final contact sheet: each on-screen number matches sources and the narration.
4. Hypotheticals are labelled on screen, and the source or as-of line is present on every data scene.
5. Footnotes and disclosures meet size and hold time.
6. Captions are burned in (or, with captions: false, the player carries them) and clearframe captions has been exported.
7. The delivery message lists the sources, what was AI-generated (voice, music, images, footage) and the spend.

• The renderer links skia-safe (MIT; Skia BSD-3), metal (MIT/Apache-2.0), rustybuzz (MIT) and kurbo (MIT/Apache-2.0). It encodes and decodes through the system FFmpeg tools (libx264 is GPL), as separate processes; review codec licensing for your distribution. Bundled Inter includes its OFL notice. Archived browser dependencies are not installed by the active package.