> ## Documentation Index
> Fetch the complete documentation index at: https://docs.manthan.systems/llms.txt
> Use this file to discover all available pages before exploring further.

# How it works

\# How Parmana Works

Parmana sits between AI systems and real-world execution to ensure that every action is explicitly authorized before it happens.

\---

\## The simple idea

AI can generate actions.

But AI should not be trusted to execute actions directly.

Parmana ensures every action is checked before execution.

\---

\## The system in 5 steps

\### 1. AI generates an action

An AI system proposes something like:

\- approve a payment

\- process a refund

\- deploy infrastructure

\- approve a claim

\---

\### 2. The action becomes a signal

The proposed action is converted into a structured signal:

\- standardized format

\- validated schema

\- traceable origin

\---

\### 3. Governance evaluates it

Parmana checks:

\- signed policies

\- deterministic rules

\- authorization conditions

It answers:

> Is this action allowed?

\---

\### 4. A decision is produced

The system produces one of:

\- APPROVE

\- REJECT

\- REQUIRE\_OVERRIDE

This decision is deterministic and reproducible.

\---

\### 5. Execution happens only if approved

Only approved decisions are passed to execution systems.

Every execution is:

\- enforced

\- tracked

\- signed

\- verifiable

\---

\## What makes this different

Traditional systems assume trust between layers.

Parmana removes trust and replaces it with verification.

\---

\## Before vs After Parmana

\### Before

```text theme={null}

AI → Direct Execution



Unverified, probabilistic, and inconsistent.



After

AI → Signals → Governance → Authorization Decision → Execution Runtime → Attestation



Deterministic, verifiable, and enforceable.



The key outcome



Parmana ensures:



AI can propose actions

but cannot execute them

unless explicitly authorized

Simple summary



Parmana answers one question:



“Was this action authorized before it was executed?”



If yes → execute

If no → block

```