AI Classifier & Agents

Fenryx ships an on-chain multi-layer perceptron (MLP) that runs once per block to classify validator behaviour. Its output drives three autonomous agents — FDM, Consensus Agent, and Routing Agent — that influence pool membership, consensus fail-safes, and tx routing in real time.

Architecture

Pools (per-block metrics)
        │
        ▼
x/classifier  ──── ONNX MLP ────► confidence score [0, 1]
        │
        ├── FDM (Fenryx Decision Module)  ──► PenaltyLevel / BAN MARK SIGN
        ├── Consensus Agent               ──► fail-safe / emergency halt
        └── Routing Agent                 ──► preferred pool for next tx

Classification thresholds

Constant	Value	Meaning
BAN_CONFIDENCE_MIN	0.80	Confidence required to issue a PermanentBan
HIGH_PENALTY_THRESHOLD	0.70	Confidence for a Heavy penalty (below: Moderate)
SIMILARITY_THRESHOLD	0.60	Zenith Pool agreement ratio below which the Consensus Agent activates the fail-safe
FAILSAFE_HALT_AFTER_BLOCKS	3	Consecutive degraded rounds (~15 s at 5 s/slot) before emergency halt

FDM (Fenryx Decision Module)

The FDM maps each validator's classifier confidence to a penalty level and decides whether to issue a BAN MARK SIGN — an ed25519-signed record broadcast over gossip.

// GET /agents/decisions
[{
  "node":        "frx1...",
  "penalty":     "PermanentBan",
  "trigger_ban": true,
  "reason":      "ByzantineVote",
  "confidence":  0.91
}]

Ban records are idempotent: re-banning an already-banned node is a no-op.

Consensus Agent

Reads the current Zenith Pool similarity ratio. If it drops below SIMILARITY_THRESHOLD:

1. The round is rolled back (IAVL commit and header construction skipped).
2. A consecutive-degraded counter increments.
3. After FAILSAFE_HALT_AFTER_BLOCKS consecutive degraded rounds the chain issues a persistent emergency halt (meta:chain_halt = true). An operator must manually clear the key and restart fenryxd.

// GET /agents/consensus
{
  "similarity_ratio": 0.94,
  "failsafe_active": false,
  "zenith_size": 21
}

Caveat: state mutations before the fail-safe check (mempool drain, tx dispatch, FDM bans) are not rolled back. Only block finalisation is prevented.

Routing Agent

// GET /agents/routing
{
  "target_pool": { "id": 2, "label": "Zenith" },
  "estimated_latency_ms": 42,
  "load_score": 0.31
}

Submit validator metrics

curl -X POST https://v2.rpc.fenryx.io/agents/metrics \
  -H 'Content-Type: application/json' \
  -d '{"node":"frx1...","cpu_pct":38.2,"latency_ms":12,"connections":44,"net_contrib":0.87}'

Endpoint reference

Method	Path	Returns
GET	/classifier/predict/:address	MLP confidence score
GET	/agents/decisions	FDM decisions from last block
GET	/agents/consensus	Consensus Agent state
GET	/agents/routing	Routing Agent recommendation
POST	/agents/metrics	Submit validator telemetry