When credit breaks, Mastermind helps you understand what matters for the CDS.

Seven layers, one of which is the model.

Promoted only when it beats its predecessor under measured gates.

Benchmark v0.1. Stages listed for versions above the promoted one are built and measured but not yet promoted; private analyses run the promoted version. As frontier models change, CDSBench reruns them and Mastermind’s analyst model can change with a new version — the customer never has to pick a vendor. Full run history, including every regression, on the model page.

A repeatable credit-event analysis process, not access to a model.

Speed

Compress the fragmented document-and-research work around one instrument into one structured analysis.

Coverage

Every objective characteristic is tested, every time — nothing depends on remembering to ask.

Historical memory

Compare the situation with structured historical CDS cases instead of starting from a blank chat.

Verification

Every material conclusion is tied back to the clause it rests on. You never have to trust the answer blindly.

Second opinion

Mastermind actively looks for reasons its own conclusion may be wrong before you see it.

Consistency

The same structure on every case, so a desk or deal team can compare analyses rather than prompts.

Uncertainty

What cannot yet be established is stated, with the document that would resolve it.

The questions a buyer asks before uploading anything.

I can just use Claude or GPT.

Yes. CDSBench shows you how Claude performs directly, question by question. Mastermind is for when you want the model wrapped in CDS-specific checks, historical precedent, evidence verification, adversarial review and structured uncertainty — and you want that measured on the same benchmark.

Is this legal advice or a determination?

No. Mastermind is decision-support software. It structures the record, runs objective checks, cites the clauses, retrieves precedent and states what is unresolved. It does not make Credit Event determinations, and its deterministic rules are provisional simplifications of the standard characteristics, shown as such.

Why should I trust your score?

Outcomes are verified against source documents by independent checks, questions are frozen in versioned benchmarks, every model receives the same record, holdout cases are never used for tuning, every failure has a public page, and errors are corrected in a public log. Methodology →

What if the models improve?

Good — CDSBench reruns them. Mastermind is built around replaceable frontier models. The benchmark tells us where generic capability has caught up and where the specialist infrastructure still adds value, version by version.

What happens to my documents?

They are parsed on the CDSBench server, stored against the email you give, never listed publicly, never added to the public benchmark and never used for precedent retrieval. CDSBench trains no models. Exactly what the current code does →