Sponsorship fit

Engineers building a sponsorship review queue use this rubric to compare a creator’s transcript with a brand brief and budget. The result supports a shortlist decision using text evidence; it does not assess video frames.

Save the rubric

Save as youtube-content.yaml:

name: Sponsorship fit
scale: { type: numeric, min: 0, max: 10, step: 0.5 }
judges:
  jev:  { type: decision, provider: typesafe, model: jev-latest }
  clef: { type: decision, provider: cloudflare, model: "@cf/cloudflare/clef-flash" }
defaultJudge: { type: ref, ref: jev }
criteria:
  - id: relevance
    name: Audience relevance to brand
    weight: 3
    inputs: [video, brief]
    anchors: { "0": Audience has nothing to do with the brand, "5": Partial overlap, "10": Audience is exactly the brand's buyer }
  - id: production
    name: Production quality (from transcript & description)
    weight: 1
    judge: { type: panel, combine: median, judges: [{ type: ref, ref: jev }, { type: ref, ref: clef }] }
  - id: safety
    name: Brand safe
    weight: 2
    scale: { type: binary, labels: [unsafe, safe] }
    anchors: { unsafe: "Profanity, political content, gambling or unsafe stunts", safe: Advertiser-friendly throughout }
    judge:
      type: panel
      combine: min
      judges:
        - { type: ref, ref: jev }
        - { type: benchmark, method: regex, pattern: "\\b(nsfw|gambling|crypto giveaway)\\b", flags: i, negate: true }
  - id: budget_fit
    name: Rate within budget
    weight: 1
    judge: { type: benchmark, method: formula, expression: "clamp(brief.budget_usd / creator.rate_usd, 0, 1)" }
aggregation:
  method: formula
  formula: "safety_n < 0.5 ? 0 : (relevance_n * 0.5 + production_n * 0.2 + budget_fit_n * 0.3)"
  normalize: [0, 100]
  passThreshold: 65

Run an evaluation

Set TYPESAFE_API_KEY, CLOUDFLARE_ACCOUNT_ID, and CLOUDFLARE_API_TOKEN. Jev scores relevance and safety; Clef-flash supplies the second production rating. The formula reads JSON inputs tagged brief and creator; relevance reads tags video and brief.

export TYPESAFE_API_KEY="your-typesafe-key"
export CLOUDFLARE_ACCOUNT_ID="your-account-id"
export CLOUDFLARE_API_TOKEN="your-workers-ai-token"
printf '%s\n' 'A transcript about budgeting for freelance designers. No sponsorship disclosures or unsafe stunts.' > transcript.txt
printf '%s\n' '{"budget_usd":2000,"audience":"freelance designers"}' > brief.json
printf '%s\n' '{"rate_usd":1500}' > creator.json
npx @verdict-ai/cli eval -r youtube-content.yaml \
  -i video=transcript.txt -i brief=brief.json -i creator=creator.json

What to change

Increase the relevance multiplier in formula and reduce another multiplier to keep their sum at 1. Raise passThreshold from 65 to 75. Replace the production panel with Jev alone to remove the Cloudflare credential requirement.

Validate your edited YAML with npx @verdict-ai/cli rubric validate youtube-content.yaml. See the rubric reference for field definitions.