Verdict · Documentation
What Verdict is
Verdict is a verification service. It runs a submitted question - optionally with your assistant's draft answer - through a structured adversarial debate between AI models from different vendors. An independent rival model critiques the draft, the models exchange rebuttals, and the debate is judged by a model from one of the two labs, alternating, on a transcript with every vendor identifier stripped before it is read. The judge returns a final answer with a confidence score and a disagreement map. All outputs are AI-generated and not professional advice. With a draft attached, your assistant's answer is the opening position and a rival model attacks it; without one, two rival models answer independently first.
Using Verdict from your AI assistant
Connect Verdict as a custom connector (Claude, Perplexity), app (ChatGPT), or MCP server (Cursor), sign in with your Verdict account, and your assistant can then call the tools listed below. Setup steps and copy-paste activation snippets are on the Connect page and, after signing in, on the Modes page of your dashboard.
The 10 tools
- debate
- Runs a full three-round adversarial debate between AI models from rival providers and returns an impartial judge's synthesis with a confidence score and a disagreement map. With caller_draft, your draft is the opening position and a rival model attacks it; without one, two rival models answer independently first. Costs 1 verdict (3 in research or complex mode or with urgency deep, 5 for an adversarial code review). Limits: question 8,000 characters, context 32,000, caller_draft 16,000; over-limit input is rejected before any model call and before any charge. Intended for verification of consequential answers, not simple lookups.
- get_debate_result
- Fetches the status and result of a previously started debate by job id. Call it about 60 seconds after debate returned a job_id. A finished debate returns the canon answer, the confidence score and the disagreement map; a failed one returns status error with a message to relay to the user as-is.
- quick_verify
- Returns a single independent critique of a provided draft answer from a rival-provider model, with an agree, disagree, or undetermined verdict. Available on paid plans and costs 0.25 verdict; on the Free plan it returns status not_available. Limits: question 8,000 characters, caller_draft 16,000. Always set caller_platform so the critique comes from a rival provider.
- set_mode
- Sets the caller's Verdict debate mode (standard, prompted_only, research, complex, full_integration, or code_adversarial). Changes only the caller's own configuration.
- get_my_mode
- Returns the caller's current Verdict debate mode and judge configuration.
- usage_left
- Returns the caller's remaining verdict quota, plan tier, the period it applies to, and the signed-in Verdict account (email, sign-in provider, mode, judge, member-since date, input limits).
- email_verdict
- Sends the finished verdict of a completed debate to the signed-in user's own email address, once per debate. Call this when the user asks to be emailed the verdict. Requires the job_id of a completed debate.
- reconnect
- Checks whether the current Verdict connection is alive and which account it is signed in as. If the session has expired, returns exact platform-specific steps the user can follow to reconnect. Call this whenever a Verdict tool fails with an authorization or session error.
- whoami
- Returns the signed-in Verdict account: email, sign-in provider, plan tier, current mode and judge setting, member-since date, and the input limits. Call it once per session before the first debate and tell the user which account you are using.
- verdict_onboarding
- Returns the activation snippet for the caller's platform and the account's current mode, exact instructions for where the user must paste it (an assistant cannot change its own settings), and the operating rules. Call it when the user wants to set up, activate or tune Verdict.
Quotas
Usage is measured in verdicts. A standard debate costs 1 verdict, Research and Complex debates 3, an adversarial code review 5, the web engine 1.5, and a quick check 0.25 (paid plans). Each plan tier has a verdict allowance per period; when it's used up, the tools return a clear message and the allowance resets automatically. Check your remaining quota anytime with the usage_left tool or on your dashboard.
| Plan | Model pair | Cost per run |
|---|---|---|
| Free | Standard pair only | Standard debate 1 |
| Pro | Standard and deep pair | Standard 1 · Research and Complex 3 |
| Power | Standard and deep pair | Standard 1 · Research and Complex 3 · Code Adversarial 5, Power only |
| Day Pass and Week Pass | Standard and deep pair | Standard 1 · Research and Complex 3 |
| Any paid plan | Web engine and quick check | The web engine 1.5 · a quick check 0.25 |
Trivial questions are refused, answered directly, and not charged. Below Power, Code Adversarial runs as a standard debate.
Privacy & data
Your questions and debates are never used to train AI models — not by us and not by the model providers. Delete your account and it's erased immediately — encrypted backups expire within 30 days. See the Privacy Policy and Terms.
Last updated: September 15, 2026.