About RedCrown

RedCrown checks whether an AI feature still meets agreed requirements, and shows the evidence behind the decision. It is built for the people who have to accept an AI delivery, review a change to it, or defend the choice later: AI consultancies delivering client features first, then internal teams with a release decision to make.

The problem we are working on

Make the AI in your product behave like software. Software has requirements written down, tests that run against them, and a record of what passed. The AI feature inside a product usually has none of those: a prompt changed, a model was swapped, the output looked fine on a few examples, and it shipped. When it fails later, nobody can say what was tested, against what, or when it stopped meeting the bar. That is the problem RedCrown exists for, and it is larger than what the product does today.

What exists today

The released product covers a narrower scope than that ambition, and this page keeps the two apart.

• Requirements as a file. A policy states the thresholds before anyone looks at the candidate: minimum cases, maximum error and ungraded rates, minimum quality, maximum regression. redcrown check reports pass, fail, or indeterminate per rule and exits with a code your pipeline can read.
• Supported comparisons and checks. Score the outputs your feature produced against your references, offline. Compare a test-harness run against the last one, check by check. Run an evaluation with RedCrown on your keys across several models. Accept an eligible run as the baseline in the app.
• Inspectable evidence. A report with a receipt for every case, the counts and the exclusions, the caveats, the version of the method, and a signed attestation. A share link the reader opens without a login. A blinded review a domain expert can cast, recorded as a review.

Not built: scheduled re-runs, alerting, monitoring, approval workflows, and root-cause tracing. RedCrown names the case that moved and the rule it failed. It does not tell you why, and it does not run unless you run it.

Why independence matters, and how ours is structured

A check is only worth something if the party running it has no stake in the answer. Most parties with an opinion about which model to run have one, and a benchmark nobody can audit is marketing. RedCrown holds the record, with no stake in which way it goes.

• No model-vendor money. RedCrown takes no payment, revenue share, or placement fees from any model provider. No vendor can buy a better result.
• No routing cut. RedCrown does not resell inference or take a margin on the calls you run. You bring your own provider keys; the calls run under your accounts, at your rates.
• Pinned, versioned methodology. Scoring runs on a fixed, versioned methodology. Every report records which version produced it, so a change in method is visible.
• Receipts, not adjectives. Every claim on a report is backed by per-case evidence and, optionally, a signed attestation you can verify offline against our published key.

Who's behind RedCrown

RedCrown was built by data scientists, not by a model lab. We spent years shipping AI systems inside other people's companies: clinical transcription, document extraction, research pipelines, forecasting. The same question came up on every project, and it almost never came from the engineer writing the code. It came from the person who had to approve the spend, sign off on the output, or defend it to a client.

"How do you know it still works?"

We never had a good answer. We had public benchmarks that ran on someone else's data, a vendor's word for it, and a few examples that looked fine. So we built the thing we kept needing: write the requirements down, check the feature against them on the client's own cases, and hand over a result they can inspect themselves, case by case, without taking our word for anything.

On names and faces. There is no founder photo on this page. RedCrown is run independently of the consultancy its team came from, and it is kept that way on purpose. That is a fair thing to ask about before you put our report in front of your client or your auditor: if you want to know exactly who you are dealing with, ask us and you will get a real person on a call and a straight answer.

For security and data handling, see the Security & Trust page. For how the product works across the app, CLI, and MCP, see the docs.

Talk to us

Evaluating RedCrown for a team, or preparing a client handoff you have to defend?