# QWED

> QWED is the open-source deterministic verification layer for LLMs and AI
> agents. It sits between a model's output and the system that would act on it,
> and it either proves the output correct with a symbolic method — SymPy, Z3,
> SQLGlot, CrossHair, JSON Schema — or refuses to let it through. Nothing here
> is scored by another model. A claim is checked against mathematics, a solver,
> a parser, or a schema, and the same input always produces the same verdict.

QWED calls this **execution governance**: a fail-closed boundary between what a
model says and what a system does. Guardrails, RAG, and fine-tuning all change
the odds that a model is right. Verification changes what happens when it is
wrong.

Every page below is also served as clean Markdown to any client that sends
`Accept: text/markdown`, which is roughly 80% fewer tokens than the HTML.

## Start here

- [QWED](https://qwedai.com/): The deterministic verification layer for LLMs and AI agents — what it checks and where it sits in a request path.
- [Why QWED](https://qwedai.com/why-qwed): The AI trust gap, and why probabilistic mitigations do not close it.
- [Execution governance](https://qwedai.com/governance): How AI systems are verified and controlled before they act, not after.
- [Vision](https://qwedai.com/vision): Verification infrastructure for production AI in finance, legal, security, and agent workflows.

## Verification engines

- [All engines](https://qwedai.com/verifiers): Index of the eleven verification engines and what each one proves.
- [Math](https://qwedai.com/verifiers/math): Calculus, matrix operations, and financial formulas checked symbolically with SymPy.
- [Logic](https://qwedai.com/verifiers/logic): Contradictions, quantifiers, bit-vectors, and policy constraints proved with the Z3 SMT solver.
- [SQL](https://qwedai.com/verifiers/sql): SQLGlot AST firewall — injection, runaway complexity, and schema violations blocked before a query runs.
- [Code](https://qwedai.com/verifiers/code): Multi-language AST analysis plus CrossHair symbolic checks; blocks eval, exec, leaked secrets, unvetted packages.
- [Schema](https://qwedai.com/verifiers/schema): Strict JSON Schema enforcement with math delegation, so computed fields are verified rather than merely well-formed.
- [Statistics](https://qwedai.com/verifiers/stats): Sandboxed pandas and polars execution with Pandera schema validation and process isolation.
- [Fact](https://qwedai.com/verifiers/fact): Claims checked against their source documents with deterministic TF-IDF and keyword analysis — no second LLM in the loop.
- [Knowledge graph](https://qwedai.com/verifiers/graph): Triple extraction and graph lookups that reject hallucinated entities and relationships.
- [Reasoning](https://qwedai.com/verifiers/reasoning): Chain-of-thought and IRAC process validation — the path to the answer, not just the answer.
- [Taint analysis](https://qwedai.com/verifiers/taint): Data-flow tracing from untrusted input to dangerous sinks in AI-generated code.
- [Vision](https://qwedai.com/verifiers/image): Image metadata analysis and multi-VLM consensus for synthetic content and unsupported visual claims.

## Open source

- [Ecosystem](https://qwedai.com/open-source): The nine core repositories, the domain packages, and the GitHub Actions, with what each one is for.
- [qwed-verification](https://github.com/QWED-AI/qwed-verification): The core engine, including AgentStateGuard for agent state governance.
- [qwed-ucp](https://github.com/QWED-AI/qwed-ucp): Unified Control Plane for deployments, configuration, and orchestration.
- [License](https://qwedai.com/license): Apache License 2.0.
- [Code of conduct](https://qwedai.com/code-of-conduct): Community standards for the open-source project.

## Learn

- [Free course](https://qwedai.com/learning): Eleven modules on neurosymbolic theory, implementation, and production patterns.
- [Documentation](https://docs.qwedai.com): Installation, API reference, and integration guides.
- [Articles](https://qwedai.com/articles): Deep dives on verification, agentic AI governance, and audit trails.
- [AI without an audit trail is a liability you can't price](https://qwedai.com/articles/ai-without-audit-trail-is-a-liability): Cascading agent failures, insurance liability gaps, and the business case for cryptographic audit trails.
- [Playground](https://qwedai.com/playground): Run math, logic, and code verification against live model output.

## About

- [About](https://qwedai.com/about): The team and why they are building deterministic verification.
- [Security](https://qwedai.com/security): Data handling, self-hosted deployment, access controls, retention, auditability.
- [Changelog](https://qwedai.com/changelog): Release history across the verification layer and its infrastructure.
- [Contact](https://qwedai.com/contact): Enterprise deployments, technical questions, security reviews, partnerships.

## Optional

- [Terms of service](https://qwedai.com/terms)
- [Privacy policy](https://qwedai.com/privacy)