QWED (Query With Evidence & Determinism) is open-source verification infrastructure for LLMs and AI agents. It checks the numbers, SQL, code and tool calls that AI produces, with solvers and parsers, before anything executes.
- A claim proven correct is admitted:
VERIFIED · ADMIT. - A claim proven wrong, one that cannot be proven, and one QWED cannot check are all denied by default.
- QWED decides; your integration, CI gate or agent runtime enforces the decision.
- Every decision leaves a record that a reviewer can check and re-run.
QWED verifies the formal statement, not what the AI meant, and it is only as sound as its engines. See what QWED does not verify, the false-accept results published at every release and the security advisories.
Engines. Proof engines check maths with SymPy, logic with Z3, SQL with SQLGlot, code with Python AST rules and structured output with JSON Schema. Advisory engines record heuristic and model-based results, never as VERIFIED. How the engine tiers work.
Core
| Repository | What it does |
|---|---|
| qwed-verification | The verification engines, API, CLI and SDKs |
| qwed-dashboard | Dashboard for a QWED server: decision ledger, records, evidence and audit chain status (work in progress) |
| docs | Source of docs.qwedai.tech |
Domains
| Repository | What it does |
|---|---|
| qwed-finance | Recomputes loans, interest, bonds, FX and risk figures exactly, and checks ISO 20022 and SWIFT messages |
| qwed-legal | Recomputes contract deadlines and liability caps, and finds contradictory clauses |
| qwed-tax | Checks AI-made tax decisions against encoded rules for the United States and India |
| qwed-infra | Proves what AWS IAM policies and network rules allow, and estimates cost before a Terraform change is deployed |
Agents and protocols
| Repository | What it does |
|---|---|
| qwed-open-responses | Checks agent tool calls and outputs before they run |
| qwed-mcp | MCP server that lets Claude Desktop, VS Code and other MCP clients run Python calculations in a restricted subprocess |
| qwed-a2a | Intercepts Agent-to-Agent messages, applies a trust boundary, checks money totals exactly and signs each decided result |
| qwed-ucp | Checks the totals, state and structure of Universal Commerce Protocol checkouts before payment |
CI and code review
| Repository | What it does |
|---|---|
| qwed-verification-action | GitHub Action that scans for secrets, unsafe code and dangerous shell scripts, with SARIF output for code scanning |
| QWED Security | GitHub App that checks every pull request with deterministic engines and signs its verdicts |
qwed-learning is a free, open course on deterministic AI verification: fail-closed design, formal verification, audit integrity and trust boundaries.
| Paper | DOI |
|---|---|
| The Layered Compression Paradox in Context Engineering |  |
QWED's current evaluation results are on qwedai.tech/evidence.
For deployments, integrations, custom guards or support, contact us or write to rahul@qwedai.tech.
- Discussions: questions and ideas
- Issues: bugs in the core
- Security: report a vulnerability privately
- Sponsor: support the work
