Product
A security layer between your application and any model.
Verity inspects every request and response, applies security and business logic, evaluates groundedness, and logs structured traces. Your application changes one line — the base URL — and keeps its SDK, prompts and provider account.
Web app, agent, or internal tool sends an OpenAI-compatible request.
Rules + trained classifier screen every message role for injection and abuse.
Your declared business rules are enforced on the request.
Only screened requests reach your provider, on your key.
The answer is graded against your context; policy is re-checked before delivery.
Every verdict is logged — inputs, outputs, signals, latency — scoped to your key.
Each layer in detail.
Detects prompt injection, jailbreak attempts and other malicious inputs. Two systems vote: deterministic rules for known attack shapes, and a Naive Bayes classifier (200 training examples, word + character n-grams) for paraphrases and obfuscation — leetspeak, homoglyphs, spacing tricks and base64 are normalized before screening. Held-out recall 88%, precision 91.7%.
Enforces your custom business rules and content restrictions, declared per request as a policy_context. Checked before the model (a prohibited ask never gets answered) and after it (a non-compliant answer never reaches the user). In live mode the policy judge fails closed.
Checks whether answers are grounded in your context or provided documents. Live mode uses an LLM judge; a content-word heuristic is the disclosed offline fallback. Unsupported claims are flagged in the trace, and policy-violating answers are blocked outright.
Logs every request, verdict and escalation for debugging and audit — stamped with a per-client identity derived from the calling key (the key itself is never stored). Retained 30 days, pruned on an ongoing basis.
Deployment options.
Create an account and call this site's gateway with your own key. Screening is production code; the model is a mock, for evaluation and integration testing. Start free
A founder-led deployment on your traffic: your provider and key, your policies, weekly reports, success criteria in writing. How pilots work
The gateway is open source (MIT). Deploy it in your own infrastructure with full control over keys, data and retention. View the repository
Implemented vs planned, honestly.
Nothing on this page is vaporware-labeled as shipped. The current state:
Input guard (rules + trained classifier, all message roles) · policy guard (request + response, fail-closed judge) · faithfulness scoring (LLM judge + heuristic fallback) · per-client traces with 30-day retention · accounts with personal API keys, rotation and revocation · free-tier monthly quota enforced by the gateway · client deployments in live mode (pilot-provisioned).
Self-serve live deployments and paid plans (Pro) · streaming response passthrough (responses currently return complete) · team workspaces and roles · configurable retention per plan · usage alerts. These ship when pilot demand validates them, in that order.
Security boundary, plainly: Verity classifies content and enforces declared policy. It is not an authorization system — permission-check sensitive actions (refunds, transfers, record changes) at your own tool and backend boundary. Full detail in the documentation.
See it work
Screen a real prompt.
The playground runs this exact pipeline against the live gateway.