> For the complete documentation index, see [llms.txt](https://docs.cournot.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.cournot.ai/architecture/fast-and-slow-paths.md).

# Fast & Slow Paths

Cournot's dual-layer architecture combines immediate execution with decentralized verification.

<figure><img src="/files/Sllk1jXk1mxNaZhpvms4" alt=""><figcaption></figcaption></figure>

### The Fast Path: Instant Finality <a href="#the-fast-path-execution-of-por-sub-second-settlement" id="the-fast-path-execution-of-por-sub-second-settlement"></a>

Anchor Nodes operate within Trusted Execution Environments (TEEs) to generate Proof of Reasoning results with sub-second settlement. The Anchor Node signs the final verdict, the reasoning root, and the committed PromptSpec hash, producing an immediately usable result.

### The Slow Path: Decentralized Security via opML <a href="#the-slow-path-verification-of-por-the-security-guarantee" id="the-slow-path-verification-of-por-the-security-guarantee"></a>

Permissionless **Sentinel Nodes** validate Fast Path results through the **opML optimistic verification** model:

1. **Evidence verification (zkTLS)**: Sentinel Nodes verify zkTLS proofs to confirm that each piece of evidence originated from an authorized source. A single invalid proof is sufficient to invalidate the entire evidence bundle.
2. **Reasoning verification (opML replay)**: If a Sentinel suspects the reasoning is flawed, it initiates an opML challenge. The frozen inputs (PromptSpec, evidence bundle, model hash) are replayed, and any divergence is isolated via the bisection protocol down to a single computation step provable on-chain.
3. **Output verification (determinism check)**: The verdict is checked against the PromptSpec's verdict schema to confirm it conforms to the allowed output space.

Because both zkTLS and opML operate under a **one-honest-node security model**, the Slow Path does not require a quorum or multi-node re-execution. A single honest Sentinel detecting a fault is sufficient to trigger correction and slashing. This makes decentralized security economically sustainable even for low-value markets.

### **Granular Challenge Mechanism**

A key innovation is Pinpoint Verification: the combination of Merkle-ized Reasoning Traces and opML bisection allows disputes to target specific reasoning steps rather than entire outcomes. A challenger does not need to re-run the full inference — they narrow the dispute to the exact leaf node where the reasoning diverged, and prove that single step on-chain.This dramatically reduces verification costs:

* **Without opML**: Dispute requires full inference re-execution across multiple nodes → cost scales with model size × node count
* **With opML**: Dispute narrows to a single computation step via bisection → cost is bounded regardless of model complexity

The result: **Instant Finality** via trusted hardware execution, with **Decentralized Security** maintained through opML's one-honest-node guarantee and economic incentives tied to stake slashing.