QWED
v7.2.2 GitHub 60

Type to search pages, engines and advisories.

    Looking for API details?Search the docs

    QWED

    Open-source verification infrastructure for LLMs and AI agents.

    QWED sits between your model and production and checks every claim with deterministic engines. Each one comes back verified,unverifiable orblocked. Nothing unproven passes.

    pip install qwedPython 3.10+

    Apache-2.0 Docker Sponsored Open Source GitHub Technology Partner
    A verified claim passes through the pangolin to production. An unverifiable claim is held for review while the pangolin half curls. A blocked claim makes the pangolin curl up and close the pipe.VERIFIED → productionUNVERIFIABLE → held for reviewBLOCKED → stoppedheld for review
    your modelproduction

    Where QWED sits#

    In your own stack, between the model and whatever acts on its output. As a library, a CLI or a self-hosted container.

    How the engine tiers work
    1. 1 · YOUR APPSends a prompt
    2. 2 · ANY LLMOpenAI, Anthropic, Gemini, NVIDIA, local
    3. 3 · QWEDRoutes each claim to an engine
    7 PROOF ENGINES · can return verified
    • MathSymPy
    • LogicZ3
    • SQLSQLGlot
    • CodeAST
    • Schema
    • StatsPandera
    • SymbolicCrossHair
    9 POLICY GUARDS · block or hold

    Shell commands, config, RAG sources, MCP tools, data exfiltration and more.

    5 ADVISORY ENGINES · signals for review

    Image, Graph, Reasoning, Consensus, Fact. Heuristic results never count as verified.

    verified → production
    unverifiable → your policy
    blocked → stopped

    Why QWED#

    Deterministic

    Claims are checked by solvers and parsers such as SymPy, Z3 and SQLGlot. The model translates; it never judges its own answer.

    Fail-closed

    Verified needs a proof. Anything that cannot be proven comes back unverifiable or blocked, never verified.

    Evidence you can keep

    Every verdict can be exported as a Verification Context v1.0 document: schema-validated and tamper-evident.

    Runs where you run

    Python, TypeScript, Go and Rust SDKs, a CLI and a Docker image. Use it with OpenAI, Anthropic, Gemini, NVIDIA or a local model.

    From install to first verdict#

    qwed init checks the engines, runs a short verification suite, and sets up your model provider.

    Full quickstart
    ~/my-project
    $ pip install qwed$ qwed init[QWED] Initializing verification engines...  [ok] SymPy    math engine ready  [ok] Z3       logic engine ready  [ok] AST      code engine ready  [ok] SQLGlot  sql engine readyRunning verification suite...  [ok] 2+2=5                    -> BLOCKED  [ok] SELECT * WHERE 1=1       -> BLOCKED  [ok] eval(user_input)         -> BLOCKEDAll engines verified. QWED is operational.
    PackageVerifiesLatestInstallLinks
    qwedVerification engines and agent security guards7.2.2pip install qwed
    qwed-financeBanking, payments and financial calculations3.0.1pip install qwed-finance
    qwed-legalDeadlines, amounts and legal claims0.5.1pip install qwed-legal
    qwed-taxTax compliance and withholding0.2.1pip install qwed-tax
    qwed-infraTerraform, IAM and cloud cost0.3.0pip install qwed-infra
    qwed-open-responsesAgent tool calls, checked before they run0.6.1pip install qwed-open-responses
    qwed-mcpMCP server for any MCP client0.2.2pip install qwed-mcp
    qwed-a2aAgent-to-agent verification0.3.0pip install qwed-a2a
    qwed-ucpUniversal Commerce Protocol transactions0.3.0pip install qwed-ucp

    Supported by

    GitHub Technology Partner Docker Sponsored Open Source Netlify Open Source plan