PROOF CENTER

Test & Benchmark Results

Controlled test, benchmark and cloud-validation results completed for Mizan products.

MIZAN CONSTITUTION v0.9

Cloud Run + Neon internal readiness

Run on a real private Cloud Run service with a Neon durable ledger.

Test	Result
Concurrent identical action	200 → 1 ALLOW / 199 VETO
Unique single-tenant	300 / 300 ALLOW
Unique multi-tenant	500 / 500 ALLOW
Cloud revision replay	PASS
Concurrency ramp	25 / 50 / 100 / 150 PASS
Rate-limited soak	5 min @ target 10 rps — PASS
Startup fail-closed	PASS
Recovery drill	PASS
Main-service 5xx scan	PASS / empty

Boundary: internal engineering/adversarial validation; not an independent pentest or production SLA.

ACTIONGATE

Action-control benchmarks

CONTROLLED SPECIFICATION BENCHMARK

ActionGate

96unique cases

100%specification accuracy

31/31unsafe baseline ALLOWs corrected

0unsafe pass-through

0safe false blocks

8/8integration & integrity checks

Controlled synthetic functional benchmark.

REALISTIC SYNTHETIC AGENT-ACTION BENCHMARK

ActionGate

1,000main requests

200adversarial holdout

100%end-to-end specification accuracy

154/154unsafe baseline ALLOWs corrected

0unsafe pass-through

0safe false blocks

Refund · payout · account change · data delete · email send

PROOFGATE / SOURCEAUDIT

HaluEval QA strict

20,000samples

93.96%accuracy

93.99%F1

94.36%hallucination recall

HaluEval QA strict-mode result.

AGENTGATE / MANDATE

Live predefined regression

32/32expected outcomes matched

ACTIONGATE / PROOFGATE

Internal final holdout

1,024cases

3,052/3,052atomic checks

163/163unsafe proposals intercepted

ACTIONGATE · STRIPE SANDBOX

Operational safety tests

84/84multi-tenant isolation

84/84credential rotation

Stripe sandbox protocol.

CYBER ACTIONGATE

Sealed synthetic holdout

200cases / 10 categories

0.0%common-surface unsafe execution

0.0%extended-surface unsafe execution

100%task recall

90.4%task precision

0deterministic / idempotence failures

Internally authored sealed synthetic holdout.