Tenzro Cortex

Tenzro Cortex is a reasoning primitive that wraps OpenMythos-style Recurrent-Depth Transformers (RDT) with Mixture-of-Experts (MoE) routing. Unlike standard token-by-token inference, Cortex exposes loop depth as a first-class, on-chain billable resource — each recurrent reasoning step is metered, priced, and cryptographically receipted.

What Is Recurrent-Depth Reasoning?

Traditional transformers perform fixed-depth forward passes. OpenMythos-family architectures introduce a Prelude → Recurrent Block → Coda structure where the recurrent block can iterate N times per token, allowing the model to allocate adaptive compute per problem.

Prelude

Initial embedding and context projection. Runs once per token.

Recurrent Block

Sparse MoE with MLA/GQA attention. Loops N times with latent-state continuation. N is bounded by the reasoning tier.

Coda

Final projection to output distribution. Runs once after the recurrent block terminates.

Reasoning Tiers

Clients select a tier when submitting a Cortex request. Each tier has bounds on loop depth and a corresponding TNZO price schedule:

pub enum ReasoningTier {
    Fast,      // 1-4 loops, cheapest
    Standard,  // 4-8 loops
    Deep,      // 8-16 loops
    Maximum,   // 16-32 loops, premium
}

Cortex Workers

A Cortex Worker is a node that advertises recurrent-depth capacity to the gossip topic tenzro/cortex/1.0.0. Workers run a Python sidecar hosting the RDT model and a Rust coordinator that meters loops, signs receipts, and routes payment.

Register a Worker (JSON-RPC)

curl -X POST https://rpc.tenzro.network -H "Content-Type: application/json" -d '{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tenzro_registerCortexWorker",
  "params": [{
    "model_id": "mythos-70b-rdt",
    "sidecar_url": "http://127.0.0.1:8765",
    "arch": "rdt-moe",
    "max_loops": 32,
    "moe_experts": 64,
    "experts_per_token": 2,
    "attn_type": "mla",
    "worker_did": "did:tenzro:machine:..."
  }]
}'

CLI

tenzro cortex register \
  --model-id mythos-70b-rdt \
  --sidecar-url http://127.0.0.1:8765 \
  --max-loops 32 \
  --moe-experts 64

tenzro cortex list
tenzro cortex reason --model-id mythos-70b-rdt --input-hex 0x...  --tier deep

Signed Receipts

Every Cortex inference produces a signed CortexReceipt containing:

loops_used — actual recurrent depth consumed
experts_activated — number of MoE experts engaged
weights_hash — SHA-256 of the served model weights
runtime_hash — SHA-256 of the runtime binary
tee_attestation (optional) — if served from a TEE
signature — Ed25519 signature by the worker DID

Receipts feed into settlement: settle_cortex_payment() routes TNZO from the client to the worker treasury with a network commission.

MCP & SDK

The cortex_reason MCP tool is available on the main Tenzro MCP server. Both the Rust and TypeScript SDKs expose CortexClient for programmatic access.

Why This Matters

Adaptive compute without Chain-of-Thought token inflation
Decentralized reasoning marketplace — providers compete on loop depth, MoE width, and TEE premium
Verifiable intelligence — every reasoning step is cryptographically attested
Agent uplift — autonomous agents can dial reasoning depth based on task difficulty and available budget

→ Tutorial: Build a Cortex-powered agent