Continuous Verification & Threat ModelingSANDBOX / TEST MODE · Razorpay Test ModeDeterministic Policy Engine Active

Deterministic Evaluation Lab & Benchmark Harness

Empirical security benchmarks (N=1,430 scenarios: 1,144 hostile prompt-injections / escalations, 286 legitimate operations) evaluated deterministically against Mandate's Policy Engine versus theoretical counterfactual baselines.

Seed:

Section A: Empirical Real Benchmark Results (Mandate Control Plane)

● Tested on FastAPI + Python PolicyEngine

Section B: Counterfactual Baseline Comparisons (Modeled Architectures)

⚠ COUNTERFACTUAL / THEORETICAL BASELINE MODELS

To illustrate why software-enforced mathematical control planes are mandatory, we compare Mandate's deterministic engine against common prompt-only and unconstrained baseline patterns.

Baseline 1

Direct LLM System Prompt Guardrails

Relying purely on natural language instructions like "Do not spend more than ₹25,000" in the system prompt.

Block Rate:34.2%

Bypass Rate:65.8%

Loss Exposure:₹14.38 Cr

Baseline 2

Static Parameter Regex Filter

Keyword and simple regex filters on tool inputs without aggregate state or delegation depth validation.

Block Rate:58.1%

Bypass Rate:41.9%

Loss Exposure:₹9.15 Cr

Mandate Control Plane

Deterministic Policy Engine (v2.0)

Synchronous 8-rule pre-flight gate, two-phase budget reservation CAS, and Zero-Gateway-Dispatch invariant.

Block Rate:100.0%

Bypass Rate:0.00%

Loss Exposure:₹0.00 (Zero Leaks)