Run a Face-Off
Paste a plain-language task, run it across a default set of top models, and get a rubric-ranked verdict with a shareable proof page.
1. Open Face-Off
Go to app.redcrown.ai/faceoff or click Face-Off in the app nav. You can also reach it from the call-to-action on the home screen.
Face-Off is available in the UI only. There is no dedicated redcrown CLI command or MCP tool for it today; the guide below covers the UI flow.
2. Paste your task
Type or paste the plain-language instruction you would send any LLM. For example: Write a Python function that merges two sorted lists. No labeled examples or ground-truth answers are required.
Face-Off uses an AI judge to score outputs, so the task can be anything a language model can respond to: code generation, summarization, rewriting, question answering, and so on.
3. Select models
A default set of top models is pre-selected: Qwen3 235B, Llama 4 Maverick, DeepSeek V4 Flash, and GPT-4o mini, all accessed via OpenRouter. Deselect any model you do not want to include.
Provider key required. Face-Off runs route through OpenRouter, so you need an OpenRouter key connected to your workspace before running. Go to Advanced then Models and keys to add one. Without a key, runs pin to the keyless demo provider (stub outputs) and the result cannot be shared as a proof.
4. Choose who judges
Select one of three options under Who judges this?
- AI only , the AI judge scores each output immediately and the ranked report is ready as soon as the run finishes.
- AI + human , the AI judge runs first, then a blind reviewer link is minted so a human expert can vote on the contested cases.
- Human only , no AI scoring; a blind reviewer link is minted and the final verdict waits for human votes.
For AI + human and human only, the existing magic-link reviewer portal is used. See Reviewer portal for details on sending and managing reviewer links.
5. Set a consistency band (optional)
Enter a repeat count to run each model N times on the same task. RedCrown reports the mean score and run-to-run standard deviation for each candidate, giving you a consistency band instead of a single-trial result. Higher N means more API calls and proportionally higher spend.
6. Contribute to the model index (optional)
Tick Help build a public model index to share anonymized win and score signals from this run. Your prompts, inputs, and outputs stay private, only the aggregate result (which model won, per-model scores, and a coarse task category) is recorded. It is off by default and opt-in per run.
7. Run and read the verdict
Click Run Face-Off. RedCrown fans the task across every selected model and the AI judge scores each output against three criteria: correctness, completeness, and follows-instructions. Each criterion is scored 1 to 10.
How the winner is picked. The report ranks candidates by score, then breaks ties by cost (cheapest that clears the quality bar wins). This is a model judgment, not ground truth. Because there is no reference answer, the AI judge can disagree with a human expert, and different runs may score differently if the model is non-deterministic. Treat the verdict as a strong signal, not a certified measurement.
To get ground-truth-backed results, bring labeled examples using Prove a task or Bring your own data.
Last verified .