Qwed Verification
Qwed Verification - AISecOps (AI Security Operations) framework for deterministic verification of AI systems. QWED verifies LLM outputs using math, logic, and symbolic execution — creating an auditable trust boundary for agentic AI systems. Not generation. Verification.
Last verified:
What is Qwed Verification?
Qwed Verification is a deterministic verification platform for LLMs and AI agents that uses formal methods, symbolic execution, and policy guards to verify AI outputs before execution or response. It acts as a trust boundary for AI systems where LLMs translate user intent into structured claims, and QWED verifies those claims with deterministic engines, providing proof-backed outcomes instead of probability-only confidence. The platform is designed for LLM verification, AI agent security, verified tool calls, prompt injection defense, and deterministic transaction verification in high-stakes workflows.
QWED provides 8 verification engines: Math (arithmetic, algebra, calculus using SymPy), Logic (propositional logic and constraints using Z3 SAT/SMT solver), Statistics (statistical claims on data), Fact (factual claims with citations), Code (security vulnerabilities via AST and symbolic checks), SQL (query safety and validity), Image (visual claim verification), and Reasoning (chain-of-thought accuracy). It includes enterprise-grade security features like SQL injection firewall, prompt injection detection, rate limiting per API key, and secure Docker sandboxing for code execution. The platform offers official SDKs for Python, TypeScript, Go, and Rust, plus a CLI for command-line verification.
QWED is designed for developers building enterprise AI systems, AI agent developers, finance and payments platforms, legal and compliance teams, and any organization working with high-stakes automation where wrong numbers can trigger payments or unsafe code can cause damage. It is particularly useful for AI reliability, verified AI agents, and scenarios requiring audit trails and evidence for incident review.
Qwed Verification pricing
Pricing model: Freemium
QWED core protocol is open-source under Apache 2.0 license, free to use, fork, and build upon. For the hosted API at https://qwed.tech/api, API keys are obtained by emailing [email protected] to request access. Early partners get founder pricing locked forever. The first 100 companies get 50% off forever. Self-hosted deployment on your own hardware is free. QWED-UCP is available on PyPI (qwed-ucp) and npm (qwed-ucp-middleware) as open-source middleware.
Qwed Verification pros
- Deterministic verification with 100% accuracy on math using symbolic solvers
- 8 specialized verification engines covering math, logic, code, SQL, statistics, facts, images, and reasoning
- Open-source core protocol under Apache 2.0 license
- Official SDKs for Python, TypeScript, Go, and Rust with stable status
- CLI tool for command-line verification of claims, logic, code files, and batch operations
- Batch verification capability to process multiple claims at once with success rate summary
- Z3 SAT/SMT solver for logic constraint verification with model generation
- SymPy-based symbolic math verification for arithmetic, algebra, and calculus
- AST-based SQL injection firewall and code security vulnerability detection
- Prompt injection defense and exfiltration checks built into SDK guards
- MCP (Model Context Protocol) security with deterministic verification and tool schema checks
- QWED-UCP middleware for AI commerce transaction verification before payment
- Fail-closed hardening with legacy CodeExecutor blocked and unknown tools default-denied
- Verification results include confidence scores (0.0 to 1.0) and verified answers
- Audit trail logging for every verification for compliance and incident review
- Async and sync client support across all SDKs with batch operations
- AgentStateGuard for deterministic structural and semantic verification of agent state payloads
- Progress-aware doom loop guard detects agents repeating actions on unchanged world state
- Production-ready infrastructure with average verification under 3 seconds (~2.8s)
- Docker sandboxing for secure code execution
Qwed Verification cons
- Requires API key obtained by emailing [email protected] (no self-service signup)
- Only 4 language SDKs available (Python, TypeScript, Go, Rust) - no Java or C#
- Logic verification requires learning QWED-Logic DSL syntax
- Code verification can timeout on complex code with loops or recursion
- Rate limits apply per API key requiring backoff implementation
- High memory usage on large batch operations without streaming
- Requires structured claims or verifiable domains - cannot verify unstructured output
- Go and Rust SDKs lack Agent API and CLI support available in Python/TypeScript
- TypeScript, Go, and Rust SDKs lack CLI support
- Ambiguous expressions are rejected in fail-closed mode requiring manual intervention
Frequently asked questions about Qwed Verification
What is QWED and what problem does it solve?
QWED (Query With Evidence & Determinism) is a deterministic verification platform for LLMs and AI agents. It solves the problem that LLMs are terrible at math, hallucinate facts, and generate unsafe code. Instead of training models to be more accurate, QWED treats LLMs as untrusted translators that convert natural language to symbolic form, then uses deterministic engines (SymPy, Z3, SQLGlot, AST) to verify the result with mathematical certainty.
How does QWED verification work?
QWED works as a trust boundary: User Query → LLM Translation → QWED Verification → Guaranteed Correct. The LLM translates user intent into structured claims, QWED verifies those claims with deterministic engines before execution or response, and you get proof-backed outcomes. For example, verifying '2+2=5' returns CORRECTED: The answer is 4, and logic constraints like '(AND (GT x 5) (LT y 10))' return SAT: {x=6, y=9}.
What verification engines does QWED provide?
QWED provides 8 verification engines: Math (arithmetic, algebra, calculus), Logic (propositional logic, constraints), Statistics (statistical claims on data), Fact (factual claims with citations), Code (security vulnerabilities), SQL (query safety and validity), Image (visual claim verification), and Reasoning (chain-of-thought accuracy). Each engine uses specialized tools like SymPy for math and Z3 SAT/SMT solver for logic.
Which programming languages does QWED support?
QWED provides official SDKs for 4 languages: Python (package: qwed), TypeScript (package: @qwed-ai/sdk), Go (package: github.com/qwed-ai/qwed-go), and Rust (package: qwed). All SDKs are stable and support sync client, async client, batch operations, and attestations. Python and TypeScript additionally support Agent API and CLI.
Is QWED open source?
Yes, the core QWED protocol is released under Apache 2.0 license and is open-source. You can use it, fork it, and build on it. The verification infrastructure is available on GitHub, and packages are available on PyPI (qwed) and npm (@qwed-ai/sdk). Enterprise features like observability, multi-tenancy, and attestations are planned for enterprise tiers.
How do I get an API key for the hosted QWED API?
To get an API key for the hosted API at https://qwed.tech/api, email [email protected] to request access. All API requests require an API key included in the Authorization header as 'Authorization: Bearer YOUR_API_KEY'. Early partners get founder pricing locked forever, and the first 100 companies get 50% off forever.
What math domains does QWED support for verification?
QWED verifies calculations across 5 mathematical domains: Arithmetic (addition, percentages, interest for financial calculators), Algebra (equations, systems for engineering tools), Statistics (mean, median, correlation for analytics dashboards), Geometry (area, volume, trigonometry for CAD tools), and Calculus (derivatives, integrals for scientific computing).
What is QWED-UCP and when should I use it?
QWED-UCP provides deterministic verification guards for the Universal Commerce Protocol (UCP), Google's open standard for AI-driven commerce. It intercepts checkout requests and mathematically verifies every calculation before payment to prevent wrong totals, bad discounts, incorrect tax, and currency errors. It includes 6 guards: Money Guard, State Guard, Schema Guard, Line Items Guard, Discount Guard, and Currency Guard. Use it when AI agents shop on behalf of users.
How fast is QWED verification?
QWED verification completes in under 3 seconds on average, with most verifications taking approximately 2.8 seconds. This is fast enough for real-time applications while being thorough enough to catch errors. The speed comes from using deterministic symbolic solvers rather than re-running LLM inference.
When should I use QWED versus prompting, RAG, or guardrails?
Use QWED when you need correctness, not just better generation quality. Prompting helps with better instructions but does not prove the answer. RAG provides better context but does not prove the conclusion. Guardrails provide better structure but do not prove semantic correctness. Human review does spot checks but does not scale. QWED provides deterministic verification but requires structured claims or verifiable domains. Use QWED when a wrong number can trigger a payment, an agent can call tools, responses must satisfy legal/policy rules, or you need evidence for audit.