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

# Bharat-Shield

> India-first AI fraud detection — what it blocks and why

## What is Bharat-Shield?

Bharat-Shield is the verification gate inside Aether Citadel. Every AI response that passes through `/v1/intent/verify` is evaluated against the original intent signature.

If the response doesn't match the intent — or matches a known attack pattern — Bharat-Shield blocks it before it reaches your users.

***

## What It Blocks

| Threat                      | How Bharat-Shield Stops It                                     |
| --------------------------- | -------------------------------------------------------------- |
| **Replay attacks**          | TTL on every intent hash — expired signatures are rejected     |
| **Intent spoofing**         | Response must cryptographically match the original intent hash |
| **Hallucination injection** | Response payload is verified against the signed context        |
| **Session hijacking**       | Intent hash is single-use — can't be reused across users       |
| **Man-in-the-middle**       | HMAC-SHA256 on the response chain — tampering detected         |

***

## The Verification Flow

```
generate_intent() → intent_hash (signed, expires in TTL)
                         │
                         ├── attached to AI request as X-Intent-Hash
                         │
              AI responds with ScaleResponse
                         │
                         ▼
             verify_response(response, original_intent_hash)
                         │
              ┌──────────┴──────────┐
              │   Bharat-Shield     │
              │                     │
         CLEARED               BLOCKED
              │                     │
       payload served         403 returned
       to user                to caller
```

***

## Verdicts

When you call `/v1/intent/verify`, you get one of two verdicts:

### Cleared

```json theme={null}
{
  "cleared": true,
  "reason": null,
  "payload": { "text": "AI output here..." }
}
```

The response is legitimate. Serve it to the user.

### Blocked

```json theme={null}
{
  "cleared": false,
  "reason": "ExpiredIntent",
  "payload": null
}
```

The response was rejected. Do not serve it. Log the reason for your own audit trail.

***

## Block Reasons

| Reason             | Meaning                                        |
| ------------------ | ---------------------------------------------- |
| `ExpiredIntent`    | The intent hash TTL has passed                 |
| `HashMismatch`     | Response doesn't match original intent         |
| `SignatureInvalid` | HMAC verification failed — potential tampering |
| `ReplayDetected`   | Intent hash was already used                   |
| `MissingFields`    | Required fields absent from the response       |

***

## India-First Design

Bharat-Shield is tuned for Indian AI workloads:

* **DPDP Act** alignment — blocks responses that would expose personal data patterns
* **Low-latency** verification optimised for India AWS/Azure regions
* **Hindi/regional language** semantic context supported in intent classification
* **UPI / fintech** intent classes (`finance`, `commerce`) with stricter thresholds

***

## What Bharat-Shield Doesn't Do

To be clear about scope:

* It does **not** moderate AI content (offensive language, toxicity)
* It does **not** rate-limit individual users (that's your app's job)
* It does **not** store the AI output anywhere — verification is stateless
* It does **not** make AI decisions — it only verifies integrity