A draft can cite a real source ID and still make a claim the source never supported. I made a small fictional YouTube-script fixture to separate mechanical checks from that harder review.

The fixture

The supplied brief describes a fictional desk kit: one recycled-cardboard tray, two cable clips and six blank labels. The tray measures 24 x 16 x 4 cm; chargers, cables and devices are not included. No productivity result, load rating or certification is supplied.

The free example files contain the fact sheet, a script, three titles, a description, a claim map and a Python 3 checker. No API key or network request is needed.

python3 audit_script.py script-source.json script-delivery.json

For this sample, the checker reports no structural errors, a 267-word script and a 47-word description. Those counts use a simple regular-expression tokenizer; counts can differ in other editors. I also checked that an unknown source ID and a 601-word script produce errors.

What can be checked mechanically?

The helper checks the pilot limits: a supplied brief of at most 1,500 words, a script of at most 600 words, exactly three non-empty titles, a description of at most 150 words, and referenced source IDs that exist in the fact sheet. It also asks for a non-empty claim map. It does not generate a script.

A deliberately small check looks like this:

for item in delivery.get("claim_map", []): for source_id in item.get("source_ids", []): if source_id not in source["facts"]: errors.append("Unknown source ID: " + source_id)

What that check cannot prove

Suppose a draft says, "This tray will double your productivity," and cites the kit-contents fact. The ID exists, so that particular mechanical check passes. The claim is still unsupported. A claim map is a review aid, not evidence of correctness.

The full helper therefore always returns semantic_review_required: true. I compare each claim to its cited fact and ask:

  • Does the source actually support this sentence?
  • Did a suggestion become a guarantee?
  • Did dimensions, quantities or exclusions change?
  • Is the CTA consistent with the brief?

The fictional sample makes no performance promise. Visual directions are creative suggestions rather than product specifications. Reading the script aloud is another separate check: a valid JSON structure says nothing about spoken pacing.

Keep the review surface small

For a short explainer, a compact source sheet and one claim map can be easier to audit than a long chat history. Keep the supplied facts, creative choices and unresolved questions visibly separate. Flag missing facts instead of filling them with plausible numbers.

Optional fixed-scope services

Disclosure: I operate the Fiverr services on my public sample and service page. It links four small pilots: Excel/CSV cleanup ($20), selectable PDF tables to Excel/CSV ($20), YouTube explainer writing ($25), and B2B content repurposing ($50). The page includes scope, delivery times and free original examples so readers can judge fit before ordering. All orders, payment and delivery stay on Fiverr.

AI generated this article and helped create the fixture and checker; I ran the structural checks locally. These are fictional demonstrations, not paid client results. A passed checker does not guarantee factual accuracy or business outcomes.