COMPARISON
Harvey, Clearbrief, and a chatbot versus the brief you received
If you are drafting the brief, Harvey and Clearbrief are the tools built for that job. GauntletScore independently verifies the document someone else put on the desk. It is not KYC, not an ID scan, not a plagiarism checker, and not another model's opinion.
Where does each tool win?
This is not a feature matrix of products we do not operate. It is the job each tool is actually for. We do not claim Harvey or Clearbrief cannot check a received file, and we do not invent capabilities for them. We say where they win, and where the job is different.
| Tool | Where it wins | The job it is not |
|---|---|---|
| Harvey | Drafting and research on work the firm is producing. | Built for the brief you are writing, not independent verification of the brief you were served. |
| Clearbrief | Cite-check of your own draft inside Microsoft Word. | Own-draft Word cite-check. That is a different job from checking the filing someone else put on the desk. |
| A generic chatbot | A fast, cheap second read of prose you already distrust. | Another model's opinion. No primary-source tool pass, no deterministic score, no signed certificate. |
| GauntletScore | Independent verification of the document you received. | Does not draft your brief, does not live in Word, and does not replace professional judgment. |
Why is own-draft cite-check a different product?
Clearbrief wins when you are still in Word, still the author, and still able to fix a bad cite before you file. Harvey wins when the firm is producing the work: drafting, research, the next version of a memo you control. Those are the right tools for that desk. The brief you were served is not that desk. You did not draft it, you cannot see which model wrote it, and you are responsible for what you do with it anyway.
Why is asking a chatbot to check the brief not verification?
A chatbot can read a PDF and tell you the cites look real. That is another model reasoning about claims. GauntletScore checks those claims against primary sources, scores them with deterministic Bayesian math that no model votes on, runs a causal and argument pass, and signs the result. The validation study is where we published what that protocol caught, and what it missed, including a failed pre-registered hypothesis and a false positive we cut from our own headline count.
What does GauntletScore not do?
It does not identify a person. It does not scan an ID. It does not detect plagiarism. It does not draft your response. It does not live inside Word. It does not replace the lawyer who has to decide what to do with the report. A claim it cannot check against a public source is reported as unverifiable, not as false.
How do you run the document you received?
Upload it at /analyze. Three free credits. No card. No call. The numbers behind the protocol, including what failed, are on /validation-study.