# Real or Render

> A pre-share check for viral video clips. Generated video is everywhere now, and so is real footage under the wrong caption. Describe what happens in a clip, paste its caption, tick the tells you actually saw on a 5x5 bingo card, and note what you have checked. The page leans one way for free; a metered read splits the caption into checkable claims, reads every tick back, lists the next checks that would settle it and drafts a reply you could send.

URL: https://real-or-render.skillsafe.ai/
API tutorial: https://real-or-render.skillsafe.ai/api.html

## Free, in the browser
- The card: 24 tells plus a free centre square, in three families - looks generated (hands, garbled signs, objects merging, faces drifting, odd physics, waxy surfaces, looping backgrounds, impossibly smooth camera, lips off, too-clean sound, very short clips, a cropped corner), wrong caption (an earlier upload, place details that don't match, the same clip under another caption, looks older than claimed) and no source (no original uploader, a repost account, no other angle, no news of a big event, no date or place, "share before it's deleted", perfect timing, a caption that tells you how to feel).
- Five corroboration stamps: who filmed it, another angle, a news report, an official or eyewitness statement, details that fit.
- The lean, by ordered non-overlapping rules: (1) a hard caption contradiction -> real but miscaptioned; (2) otherwise two or more generation tells -> likely generated; (3) otherwise two or more corroborations -> likely real; (4) otherwise not enough yet.
- Bingo lines (rows, columns, diagonals, centre free), a family tally, warnings, Markdown and JSON export, saved checks, the next checks added to your notes as lines to fill in before a second read, the reply copied or shared.

## Metered (gpt-terra), one task: check
Reply: verdict (real_but_miscaptioned / likely_generated / likely_real / not_enough_yet), title, summary, claims [{claim, status backed|contradicted|unchecked, basis}], tells [{tell, reads_as generated|miscaptioned|unsourced, why}] one per tick, checks [{step, how, settles}] 3-5, reply.
The model never sees the clip. It may mark a claim backed or contradicted only when the person's notes or card report something; it names no products or websites, accuses no one, and never claims to have searched.

## Limits
The tells are rules of thumb, not a detector: real phone video can trip them and good generated video can dodge them. Treat any read as a reason to check, not a ruling.

## Source
Built on the SkillSafe skill @jwynia/claim-investigation - https://skillsafe.ai/skill/@jwynia/claim-investigation