# How AI Settlement Works

## Settlement

Every market on Delphi is settled by an AI model. This page explains those two settlement paths: **\[1]** non-verifiable and **\[2]** verifiable, and what "verifiable" actually *means* in the context of Delphi.

### The Two Settlement Paths

When creating a market, the creator selects an AI model that will serve as the judge. There are two categories of models available, and the choice determines how the market's resolution can be trusted.

#### Non-Verifiable Models

Non-verifiable models are cloud-hosted LLMs like GPT-5.3, Gemini 3, and Claude 4.6. They have tool access, which means they can look up data sources on the internet directly as part of the settlement process.

The trade-off is reproducibility. 

Because these models run on third-party infrastructure and may behave non-deterministically, someone else can't re-run the exact same settlement and guarantee a bitwise identical result. You're *trusting* that the model and the creator handled the process correctly, but there's no cryptographic proof.

#### Verifiable Models

Verifiable models are local models that run inside [Gensyn's REE (Reproducible Execution Environment).](https://app.gitbook.com/s/jHECdpSAZDuPfU2oZmM2/) 

When a creator settles a market using a verifiable model, the process works like this:

1. The creator *manually retrieves* the relevant data from the specified sources (e.g., checks the precipitation level using Weather Underground, such as in [this example](https://docs.gensyn.ai/testnet/create-a-market#settlement-prompt)).
2. The creator appends that data to the settlement prompt.
3. The combined prompt is run through the model inside REE.
4. REE produces a [settlement receipt](https://app.gitbook.com/s/jHECdpSAZDuPfU2oZmM2/ree/receipts) which is a record containing the model used, the full prompt (including the appended data), and a cryptographic hash of the computation.

Anyone can then take that receipt and [run the same computation on their own machine](https://app.gitbook.com/s/jHECdpSAZDuPfU2oZmM2/ree/examples#verifying-a-receipt) in reproducible mode. If the hash matches, they've independently verified that the settlement was computed correctly.

{% hint style="info" %}
Markets that use verifiable settlement are tagged with a *Verifiable* label on their market card and within the market detail view.
{% endhint %}

### What Verification Proves

Verifiable settlement proves that *the computation was done correctly given the inputs.* It does not independently confirm whether the data the creator provided is factually accurate.

For example, say a market asks: *"Will the price of oil exceed $120 by April 14?"* At settlement time, the creator checks the price, finds it's $107, and includes that data in the settlement prompt. The verifiable model evaluates the prompt and data, and returns "No." Anyone can verify via the receipt that the model correctly concluded "No" based on a price of $107.

But if the creator had entered a different price, such as by mistake, the model would have computed against that different number. The verification would still confirm the computation was correct *for the data provided.* It wouldn't catch that the data was wrong.

So what verification gives you is computational integrity: a guarantee that the AI model processed the settlement prompt and data honestly and reproducibly, without tampering. Combined with transparent data sources (which are listed on every market), this creates a system where settlement can be independently audited.

Non-verifiable models skip this layer entirely and instead you're relying on trust in the cloud provider and the creator. That's fine for many use cases, but verifiable markets provide a stronger assurance.

### How to Evaluate a Market's Settlement

When you're deciding whether to trade on a market, the AI Settlement section on the market page gives you everything you need to assess how it will be resolved:

* **Model:** Which AI model will settle the market, and whether it's verifiable.
* **Settlement Prompt:** The exact prompt the creator will use. You can evaluate whether it's clear, specific, and likely to produce an unambiguous result.
* **Data Sources:** Where the creator says they'll get their data. For verifiable markets, the creator will need to manually include this data at settlement time.

The *Creator Stats* section also helps you gauge trustworthiness: how many markets they've created, how many they've settled, and their settlement percentage. 

{% hint style="success" %}
A creator with a 100% settlement rate who writes clear prompts and uses verifiable models is generally a reputable and trustworthy market creator, so look for high settlement rates.
{% endhint %}