> For the complete documentation index, see [llms.txt](https://docs.verio.network/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.verio.network/vision.md).

# Vision

From trusting providers to checking evidence.

## The evidence gap

AI has reached the point where the question is no longer whether models are available, but whether claims about them can be trusted. Open-source AI has solved distribution: the Hugging Face Hub alone lists more than three million models, with standard repositories, versioning, model cards, and inference integrations. Availability, however, has outpaced accountability.

Model cards describe training data, intended use, limitations, and evaluation results, and they make models legible to other developers. They do not turn a statement about a model into evidence that the statement is correct, so every claim still rests on trust in whoever made it.

Production AI makes this harder. A single output may depend on a base checkpoint, several fine-tunes, the datasets they were trained on, a tokenizer, a prompt template, an inference runtime, decoding parameters, and an execution provider. A service that reports which model it ran, a model page that reports a benchmark score, a derivative that declares its ancestry, and a developer that describes its training data as licensed all make claims that an independent party cannot easily verify.

These unverifiable claims now carry real consequences. Routers select among providers automatically, agents pay for and act on model outputs, and regulators require providers to document how their models were built. Once a claim determines who is paid, which transaction executes, or whether an obligation is met, the provider's own account of what happened is no longer sufficient. Rights holders have already filed 145 copyright suits against AI companies in the United States alone, over both the works used in training and the content models generate.

The industry's answer so far has been more documentation, self-reported metrics, and occasional outside audits. None of these changes who must be trusted, or shows what a model actually learned from.

## What Verio changes

Verio replaces "trust the provider" with portable evidence that anyone can check. The evidence layer sits on top of existing AI infrastructure rather than replacing it, and records for every consequential claim:

* which artifact the claim concerns,
* who made it,
* how strongly it has been checked, and
* which assumptions remain.

Three primitives deliver this. An [Artifact Passport](/protocol/artifact-passports.md) gives each model, dataset, adapter, evaluation, and execution profile a stable cryptographic identity. An [Inference Receipt](/protocol/inference-receipts.md) binds a model invocation to the exact artifact version, execution configuration, provider, and result. The [Provenance Graph](/protocol/provenance-graph.md) connects derivative artifacts to their upstream dependencies for attribution, reputation, and settlement.

## Verification without migration

Other approaches ask developers to move workloads onto new networks or specialized hardware. Verio lets a provider become verifiable simply by emitting receipts and accepting replay. The blockchain coordinates shared state; model weights, datasets, and computation remain off-chain.

The technical design is intentionally narrow. Verification is built from signatures, content commitments, deterministic execution profiles, independent replay, redundant execution, randomized audits, and open disputes. This keeps the trust model understandable and places the strongest guarantees on claims that can actually be reproduced or challenged.

## Evidence as an economy

When receipts record which artifact produced a result, and the provenance graph records where that artifact and its training data came from, payment can flow to the creators and contributors a result depended on, and false claims carry financial consequences. Verio turns verification into a market in which evidence, and fair credit for the work AI learns from, become default properties of AI outputs.