Vairify — the AI Trust Layer. Every response verified against evidence, with cryptographic proof — before your users ever see it. One URL change. No changes to your code.
No credit card·Deploy in 5 min·One URL change
The difference between "the system told you it might be wrong" and "the system fixed it before anyone saw it". Every response passes through the full trust pipeline — automatically.
Set the bar per environment. Vairify decides what happens to every response before delivery.
Pass-through — transparent telemetry only.
Flag risky responses, deliver with probe annotation.
High-risk claims never reach the user — policy violation returned instead.
Repair the response before delivery. User sees the corrected answer.
Consensus across models when certainty is low.
Critical-domain answers queue for approval.
Full decision trace + evidence on every delivered answer.
Roadmap (V2): nothing reaches the user until it passes your quality policy. Today V1 detects, verifies and blocks — repair ships with the V2 gateway.
The business promise
"Flip one mode on and every end-user answer meets your quality policy — automatically. No dashboards to watch, no manual review of every response."
Every confident answer is a statistical guess. In production, one fabricated claim can reach a customer, a report, or a regulator — and you find out after it costs you.
Chatbot invented a bereavement-fare policy. Court ruled the airline must honour the hallucination. The company argued the bot was "responsible for its own actions" — the tribunal disagreed.
src: BC Civil Resolution Tribunal · src: CNBCA produced AI tool fabricated legal and financial citations. The error surfaced only after a client paid for the flawed output. Damage: reputation + AUD 440k class action.
src: Reuters · src: ABC NewsYou can't make the model omniscient. You can make its output provable.
Post-hoc. LangSmith, Promptfoo, W&B report scores after the fact — in a dashboard nobody watches during runtime. No enforcement, no proof.
Patterns, not truth. Lakera, Guardrails AI scan for words and jailbreaks. A confidently false but clean-sounding claim passes untouched.
No guarantee. OpenAI, Anthropic, Google don't certify factual correctness — and their ToS says so. Legal exposure stays with you.
Runtime evidence. Intercepts the response, verifies the claims, attaches a verdict and a cryptographic proof — before the user sees it.
If your AI starts answering nonsense tomorrow, Vairify catches it before your users see it. Nine engines, six verdict classes, cryptographic proof — the peace of mind comes built in.
7-signal detector: vague quantifiers, false certainty, unverifiable claims, self-promotion.
Flags "100%", "guaranteed", "infallible" — precision without evidence.
Detects rhetorical performance dressed as fact.
Policy checks, blast radius, firewall rules per organization.
Extracts verifiable claims and attaches supporting evidence from sources.
Measures how destructive a failing claim would be in your stack.
4-dimension score with governance weight — per claim, per model, per org.
Reproduce any past verdict exactly. Root-cause in minutes, not days.
Catches answers shaped to please rather than to be true.
6 verdict classes
VERIFIED · PARTIALLY_VERIFIED · UNVERIFIED · HALLUCINATED · CONTRADICTED · PENDING — each with confidence, evidence chain, and proof hash.
Paste a sentence an LLM might produce. Vairify classifies it — either against the live API or the same detection patterns locally.
Your app sends the prompt to the LLM — through Vairify's OpenAI-compatible endpoint.
Vairify pulls atomic operational claims out of the response: facts, numbers, dates, forecasts.
9 engines score each claim against evidence and policy. Deterministic validators check numbers and dates.
VERIFIED → CONTRADICTED, with confidence, evidence chain, and SHA-256 proof hash.
Not "a history of verdicts" — a command center for AI response quality.
Docker image, single binary, zero outbound data except the verification you choose to report. Runs behind your firewall.
Memory-safe, async, single static binary. No runtime dependencies, no supply-chain surprises.
Every verdict is a SHA-256-anchored evidence chain. Replayable end-to-end. Audit-ready by construction.
HMAC-SHA256 machine keys. Rate-limited auth routes. Per-org governance and trust scoring.
SOC 2 in progress · SSO/SAML on roadmap — we're honest about where we are. Self-host covers the gap today.
Your prompts and responses are not used to train anything.
Zero cookies on the landing. No behavioral fingerprinting.
Self-host mode: verification runs inside your network. Telemetry is opt-in.
Evidence and verdicts live in your database. You own the audit trail.
No credit card required·Self-host option·Cancel anytime
Atomic operational claims inside LLM responses — facts, numbers, dates, forecasts, instructions. Each gets one of six verdicts with confidence and evidence. We don't judge style; we judge provability.
Dashboards score after the fact and don't change runtime. Vairify sits on the request path and returns a verdict with every response — plus a cryptographic proof you can show an auditor.
Point your client at the Vairify endpoint (OpenAI-compatible). That's the integration. Most teams verify their first claim in under 5 minutes. Self-host: one Docker container.
In self-host mode, no — verification runs in your perimeter. In cloud mode, prompts and verdicts are stored in your tenant database and never used for training. Telemetry is opt-in.
Any OpenAI-compatible endpoint: OpenAI, Anthropic, Gemini, Ollama, vLLM, self-hosted models. The proxy doesn't care who generates the text.
SOC 2 is in progress. Today, self-host + cryptographic evidence chain + replay covers the audit-readiness gap. We're transparent about certifications instead of faking badges.
No — and anyone who promises that is lying. Detection is pattern + evidence based, with deterministic numeric validation. We reduce the risk dramatically and give you proof of what passed. That's the honest position.
A browser copilot for AI safety. Get instant warnings before you paste sensitive data, catch answers that cite nothing, and flag risky queries — right inside ChatGPT, Claude, Gemini and DeepSeek.
We build it only if you want it. No spam, no noise.
No credit card·Deploy in 5 min·Self-host option