> 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/overview/the-oracle-evolution/oracle-3.0-the-intelligence-oracle/opml-making-ai-inference-provable.md).

# opML: Making AI Inference Provable

opML applies the optimistic verification pattern, proven at scale by optimistic rollups, to AI inference. The core insight: rather than requiring every node to re-execute the same inference (which is prohibitively expensive for large models), opML assumes the result is correct and provides a mechanism for any single honest node to prove it wrong.

**How opML works in Cournot**:

1. **Execute once**: An Anchor Node runs the AI inference inside a TEE and publishes the result along with the frozen inputs (PromptSpec, evidence bundle, model hash, execution parameters).
2. **Challenge window**: The result enters a challenge period during which any Sentinel Node can dispute it.
3. **Deterministic replay**: If challenged, the inference is replayed with the same frozen inputs, fixed model weights, and controlled execution environment. Because Cournot freezes all inputs at the PromptSpec and evidence stages, the replay target is fully specified.
4. **Bisection protocol**: Rather than replaying the entire inference, the challenger and defender engage in an interactive bisection protocol, narrowing down to the exact computation step where results diverge, then proving that single step on-chain.

**The result**: Verification cost is reduced from "run the full model N times" to "one honest node submits a compact fraud proof." This is what makes AI verification economically viable at scale.

#### **The zkTLS + opML Verification Stack**

Together, zkTLS and opML form a complete verification stack where **a single honest node can prove the entire pipeline**:

<figure><img src="https://2241883016-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FWPGzY99JooaGhfRwwXbw%2Fuploads%2FEhYoKkRscleJqJWeTsWW%2F4..%20ZKTLS%20OPML.png?alt=media&amp;token=17728da4-407e-472b-95bc-33966ff64558" alt=""><figcaption></figcaption></figure>

**This one-honest-node assumption** is what separates Cournot from architectures that require multiple nodes to independently run the same expensive inference to reach consensus. Those approaches make the dominant cost, model inference, multiplicative with the number of nodes. Cournot runs inference once and makes verification cheap.