Current status: public beta / controlled-pilot technology. Production use requires separate review.
CONTROLLED EVIDENCE

Show what was tested. Keep the limitation next to the number.

This page contains the evidence we are comfortable using as product proof today. It intentionally excludes incomplete research lines, weak benchmark pockets and results that would create a misleading flagship claim.

ACTIONGATE / PROOFGATE

Internal final adversarial holdout

1,024cases
3,052/3,052atomic checks
163/163unsafe proposals intercepted
0unsafe pass-through
Scope: internal controlled final holdout. No real payment API call or money movement. Not an external benchmark.
STRIPE SANDBOX

Operational control protocols

84/84multi-tenant isolation controls3 tenants; controlled Sandbox protocol
84/84credential rotation controlscontrolled dual-slot rollover protocol
0cross-tenant leaks in named protocol
0secret exposure in named rotation protocol
Scope: controlled Stripe Sandbox validation. Not an independent pentest, production IAM certification, HSM/KMS certification or live-payment attestation.
SOURCE VERIFICATION

HaluEval QA full · strict mode

20,000samples
93.96%accuracy
93.99%F1
94.36%hallucination recall
Scope: the documented HaluEval QA full strict-mode run only. Dataset-specific benchmark evidence is not a universal correctness guarantee.
AGENTGATE

Live predefined regression

32/32expected outcomesrecorded live suite
ALLOWsafe read path
VETOscope / intent / contract violations
REVIEWmissing evidence / approval
Scope: internal live regression over the predefined cases. This is not a security certification for arbitrary agents, tools or deployments.

What is deliberately not promoted here

Research / benchmark lines such as FiFAR, ULB, Cognitive Gate and other exploratory datasets remain R&D evidence rather than customer SKUs. Benchmark pockets that are weaker or incomplete are not converted into broad marketing claims.

What would raise the evidence level

Independent security assessment, paid design-partner pilots, customer-labelled Shadow Mode results, production references and repeatable commercial outcomes would move Mizan from controlled technical evidence toward external validation.