Verdict | AI Output Moderation Laravel Script With Mobile App
Original price was: $29.00.$5.00Current price is: $5.00.
Your AI agent just told a customer their refund was approved. Was it?
Verdict checks what an AI agent produced before it reaches anyone or triggers anything, and returns a ruling your code can branch on: pass, flag, or block, with the evidence behind it.
One API call per output. No second chat model grading the first one, and no paragraph of prose to re-parse into a boolean.
Key features
- One typed ruling per output. Post the agent output plus the context it had, get back pass, flag or block with a confidence score your code can branch on.
- Eight independent checks, each switched on and thresholded per policy: hallucination, contradiction, policy violation, PII leak, high risk action, code convention, tone and brand safety, routing.
- Worst ruling wins. Checks never average out. One block is a block, so a serious failure cannot be diluted by seven passes.
- Unanswerable means flag, never pass. A model timeout fills the review queue instead of letting unverified output through.
- Full evidence trail. Every case stores the per check answer, confidence, policy version and who ruled on it, so an audit can reconstruct any decision.
- Policies per agent source. Each calling app gets its own API key, its own policy and its own thresholds.
- Human review workflow on web and mobile, with a reviewer role that can rule on cases and reach no configuration at all.
- Filament 4 admin with an analytics overview: rulings per day, which checks are doing the work, and tone drift over time.
- Webhooks on flagged and blocked rulings, signed with HMAC SHA256 so the receiver can prove the call came from your install, and retried with backoff when you run a queue worker.
- Flutter app for iOS and Android at full parity with the reviewer dashboard, including push notifications.
- White label throughout. Name, logo, support address and palette are editable from the admin and reach web, app and email with no rebuild.
- Accessible by construction. The branding screen rejects a palette whose pass, flag and block colors cannot be told apart under simulated color blindness.
- Four step web installer, from unzipped folder to working install with no file editing, then it closes itself.
- Sixteen email templates, custom error pages, a live chat widget, and a thirteen page illustrated manual, also supplied as a 44 page PDF.
- Tested. 298 automated tests covering the engine, the API, the gates, the installer and the admin.

There are no reviews yet.