Source verification
Source verification: every figure in your script, checked against its source
CLAQON source verification starts before writing: it gathers citable sources or uses yours, and the script can only rely on verbatim excerpts. Then a deterministic check, HARD QC, rejects figures and absolutes that are not in the evidence. Anything that does not pass never reaches the voice or the visuals.
- Only verbatim excerpts from the source
- A deterministic check before spending
- Human review when needed
What it solves
Why fact-check a short before publishing it?
An AI model fluently writes sweeping statements (“always”, “never”, “the best”) and figures that sound plausible. On an informational channel, a made-up fact costs credibility, and fixing it after publishing is expensive. Asking the model to “be careful” is not enough: it has to be checked.
CLAQON separates the two tasks. The model writes, but the decision on whether a script moves on is made by a deterministic check that compares each claim with the approved evidence. And it does so before spending on voice, images or video.
How it works
From the search for sources to the approved script
Research
It searches for citable sources with Tavily or uses the URLs you paste. Your sources are downloaded through the safe outbound network, up to 2 MB each, and the script cites your links with no search cost.
Compile the evidence
Each claim needs an excerpt that appears word for word in the source, is on topic and is of valid length. Anything that does not comply is rejected and can no longer be cited.
Score the opportunity
A deterministic score decides whether the topic deserves the piece: it passes at 60 or above and penalises topics that are too short or too long and “easy money”, casino or betting angles.
Write with citations
The AI model you choose writes scenes, hook, titles and descriptions. Each claim must carry an explicit citation to the approved evidence.
Pass HARD QC
The script only moves on at 100/100. If it fails, the flagged issues are repaired; if it still does not pass after the attempts, it waits for your review before the voice and the visuals.
The check
What does HARD QC check?
It is a deterministic quality check: the same rules for every script, with no other AI giving its opinion. Among other things, it checks:
- Absolutes and superlatives not supported by the evidence
- Figures that are not in the approved evidence
- Rejected claims or claims with no declared source
- Assertive sentences with no evidence to support them
- Bands for word count, length and number of scenes
- That the piece opens with a hook and closes with a call to action
- Repeated scenes
- That there is a supported title
- Promised resources that do not exist
- Rights and suitability declaration
- Origin of the music
- Declared monetisation
- That the response is valid JSON
Capabilities
Research, repair and diagnosis
- Your own sources
- You paste the URLs in Production and click “Check links”. The workflow uses that material, skips the paid search and cites your links.
- Verbatim evidence
- An excerpt that does not appear word for word in the source, is off topic or is of invalid length is rejected.
- Targeted repair
- It fixes only the flagged errors, keeps the narration and its order, does not relax the word, length and scene bands on the last attempt and does not invent citations.
- Human review only when needed
- A script that does not pass after the attempts is held before the voice and the visuals. Once you approve it, the rest of the process continues on its own.
- The “Why they did not complete” catalogue
- 65 causes, each with its stage, its explanation and its fix (32 of them from the check stage). It is extended without redeploying the workflows, and old records are read again with the new table.
- Honest metrics per model
- “Quality success rate”, “Operational success rate” and first attempt versus repair, always with their denominators. “No data” does not mean 0%.
Limits
What verification cannot promise
- Rigour has a cost: a script that finds no support is not published, and there may be slots in which the account publishes nothing.
- Research is only as good as its sources: the ones Tavily finds (paid, with your key) or the ones you provide.
- The check reviews the text of the script. Stock clips are chosen and scored automatically, and that is not the same as a human visual review.
- The success metrics are those of your workflow and your tasks, not a universal property of each model.
Frequently asked questions
Questions about source verification
What is HARD QC?
It is the deterministic quality check for the script. It requires 100/100: no errors and sufficient evidence coverage. Among other things, it checks that every figure and every absolute is in the approved evidence, the length and scene bands, the hook and the call to action, and that the title is supported.
Does CLAQON guarantee that my videos are error-free?
No. It rejects figures and claims without evidence and shows you why, but it cannot guarantee the absence of errors: a source may be wrong. What the process does ensure is that every script goes through the same checks and that none reaches the voice without passing them or without your approval.
Can I use my own sources instead of a search?
Yes. You paste your URLs and click “Check links”: the workflow uses that material, does not spend on search and cites your links. If you do not provide sources, research uses Tavily with your key.
What happens when a script does not pass the check?
A targeted repair is attempted that fixes only what was flagged, without relaxing the bands or inventing citations. If it still does not pass after the attempts, the script waits for your review and nothing is spent on voice or visuals until then.
How do I know why a piece was not published?
The diagnosis explains it with one of its 65 documented causes: the stage where it happened, the reason and what needs changing. See the control room.
Which AI model gets the most scripts approved?
It depends on your account and your topics. For each model, the panel shows the “Quality success rate”, the “Operational success rate” and first attempt versus repair, with their denominators, so you can compare them with your own data. See AI models.
Early access
Produce with sources you can show
If your channel lives on facts, tell us your niche and how you verify today. We will tell you whether CLAQON fits the way you work.