stemedb/ai-lookup/features/admission-control.md

# Admission Control (The Shield)

Phase 7A introduces graduated proof-of-work admission control to defend against spam, Sybil attacks, and knowledge poisoning when Episteme is open to millions of agents.

## Overview

New or untrusted agents must solve BLAKE3-based proof-of-work puzzles before their assertions are accepted. As agents demonstrate good behavior, they graduate to higher trust tiers with reduced PoW requirements and increased quotas.

## Key Concepts

### Trust Tiers

| Trust Range | Tier       | Quota Multiplier | PoW Required |
|-------------|------------|------------------|--------------|
| 0.0-0.3     | Untrusted  | 0.1x (1,000/hr)  | Yes          |
| 0.3-0.5     | Limited    | 0.5x (5,000/hr)  | Yes          |
| 0.5-0.7     | Verified   | 1.0x (10,000/hr) | No           |
| 0.7-0.9     | Trusted    | 2.0x (20,000/hr) | No           |
| 0.9-1.0     | Authority  | 10.0x (100k/hr)  | No           |

### PoW Graduation

| Assertions | Trust Score | Difficulty | Approximate Effort |
|------------|-------------|------------|-------------------|
| 0-9        | < 0.6       | 16 bits    | ~16 seconds       |
| 10-49      | < 0.6       | 1 bit      | Trivial           |
| 50+        | any         | 0 bits     | Exempt            |
| any        | >= 0.6      | 0 bits     | Exempt            |

## HTTP API

### GET /v1/admission/status

Get admission status for an agent.

**Query Parameters:**
- `agent_id` (required): Agent's Ed25519 public key (hex, 64 chars)

**Response:**
```json
{
  "agent_id": "abc123...",
  "tier": "Verified",
  "trust_score": 0.55,
  "assertions_count": 42,
  "pow_difficulty": 0,
  "pow_required": false,
  "base_quota_limit": 10000,
  "effective_quota_limit": 10000,
  "quota_multiplier": 1.0,
  "assertions_until_reduced_difficulty": null,
  "assertions_until_exemption": null
}
```

### Request Flow

1. Agent submits assertion with `X-Agent-Id` header
2. AdmissionLayer checks trust tier and assertion count
3. If PoW required:
   - Return HTTP 428 with difficulty in response body
   - Agent solves `BLAKE3(nonce || agent_id || timestamp)` with required leading zeros
   - Agent resubmits with `X-PoW-Nonce` and `X-PoW-Timestamp` headers
4. PoW verified, request proceeds to MeterLayer
5. On success, assertion count increments

### HTTP 428 Response (PoW Required)

```json
{
  "error": "Proof-of-Work required",
  "code": "POW_REQUIRED",
  "required_difficulty": 16,
  "pow_required": true,
  "agent_assertions": 3,
  "agent_trust_score": 0.5
}
```

## Headers

### Request Headers

| Header | Description |
|--------|-------------|
| `X-Agent-Id` | Agent's Ed25519 public key (hex, 64 chars) |
| `X-PoW-Nonce` | PoW solution nonce (decimal) |
| `X-PoW-Timestamp` | PoW solution timestamp (Unix seconds) |

### Response Headers

| Header | Description |
|--------|-------------|
| `X-Trust-Tier` | Agent's trust tier name |
| `X-PoW-Required` | "true" or "false" |
| `X-PoW-Difficulty` | Required difficulty in bits |
| `X-Quota-Multiplier` | Tier quota multiplier |

## Implementation Details

### Core Types

**TrustTier** (`stemedb-core/src/types/trust_tier.rs`):
- Enum with 5 tiers: Untrusted, Limited, Verified, Trusted, Authority
- Methods: `from_score()`, `quota_multiplier()`, `requires_pow()`

**PowProof** (`stemedb-core/src/types/pow.rs`):
- Struct with `nonce`, `agent_id`, `timestamp`
- Methods: `verify()`, `compute_hash()`, `leading_zeros()`, `solve()`

**AdmissionConfig** (`stemedb-core/src/types/pow.rs`):
- Configurable thresholds and difficulties
- Default: 16-bit initial, 10 assertions reduced, 50 graduated

### Storage

**AdmissionStore** (`stemedb-storage/src/admission_store/`):
- Wraps TrustRankStore (reuses existing trust score + assertion count)
- No new storage keys needed
- Methods: `get_admission_status()`, `verify_pow()`, `record_assertion()`

### Middleware

**AdmissionLayer** (`stemedb-api/src/middleware/admission.rs`):
- Tower middleware applied before MeterLayer
- Extracts PoW headers, verifies proofs, returns 428 on failure
- Bypasses health checks and admission status endpoint

## Client Implementation Guide

To solve a PoW puzzle:

```rust
use stemedb_core::types::PowProof;

// Get your agent's Ed25519 public key
let agent_id: [u8; 32] = ...;
let timestamp = SystemTime::now().duration_since(UNIX_EPOCH)?.as_secs();
let difficulty = 16; // From 428 response

// Brute-force search for valid nonce
let proof = PowProof::solve(agent_id, timestamp, difficulty);

// Submit with headers
// X-Agent-Id: {hex(agent_id)}
// X-PoW-Nonce: {proof.nonce}
// X-PoW-Timestamp: {proof.timestamp}
```

## Router Functions

Three router variants are available:

1. `create_router()` - No admission control or metering
2. `create_router_with_meter()` - Metering only (quotas)
3. `create_router_with_admission()` - Full protection (PoW + quotas)

## Security Properties

- **Replay Prevention**: Proofs expire after 5 minutes
- **Agent Binding**: Proof includes agent_id, cannot be reused by others
- **Asymmetric Cost**: O(2^difficulty) to solve, O(1) to verify
- **Fail Open**: On store errors, requests are allowed (availability > strictness)
- **Defense in Depth**: API layer primary, ingestion layer secondary

## Future: Phase 7B (EigenTrust)

Phase 7B will build on this foundation:
- Peer-to-peer trust propagation (trust agents you trust)
- Network-wide reputation scores
- Dynamic tier adjustments based on global consensus