Ask a natural-language question. Provenance answers from the records you control — citations, evidence, and stance — and hands that grounded context to your LLM via API or MCP. Your LLM composes the final answer using only what the records actually say. Same engine also catches hallucinations in existing LLM output, claim by claim.
Most "AI grounding" tools charge you per token, per query, or per seat — and lock you to their LLM, their vector store, their pricing whim. Provenance is a flat per-machine license that runs on your hardware. Three things stop scaling against you the moment you install it.
OpenAI on Monday. Anthropic on Tuesday. Local llama on Thursday. Provenance hands the same grounded prompt to whatever you point it at — no embedding model lock-in, no preferred-vendor tax, no migration cost when you switch.
The records never leave your install. Validiti never sees the question, the records, or the answer. Pharma sensitivity, attorney-client privilege, regulated-industry data sovereignty — all stay where they need to.
Flat per-machine, per-month. Verify a thousand claims or a million — same price. No per-token meter, no per-query gotcha, no surprise overage. Predictable line item, indistinguishable from any other software seat.
Same engine, two flows. Start your workflow at either end.
You ask Provenance a question in plain language. Provenance returns the verified evidence from your records — citations, stance, gaps. That grounded context becomes the prompt for the LLM you choose. Your LLM never invents what isn't in the records.
PRIMARY FLOW
Already have an LLM draft? Paste it in. Provenance returns the same content with every claim labeled — VERIFIED, PARTIAL, NO SOURCE — checked against the same records. Catches hallucinated drug names, fictitious citations, unsupported facts.
SECONDARY FLOW
Six things every other "grounded LLM" approach pretends to do — Provenance actually does, in milliseconds, on your own hardware.
Ask a natural-language question. Provenance searches the records you registered, returns the actual evidence — citations, stance, gaps in coverage — and a structured prompt your LLM can consume. The answer is grounded before the LLM sees it.
API or MCP. Validiti Accelerate, OpenAI, Anthropic, or a local model — your call. Provenance returns the verified context; your LLM uses its full context window to compose the final answer. We never lock you to one model.
Provenance runs locally. Point it at the records you trust — pharmacovigilance archives, your case-law set, your internal documentation, your peer-reviewed literature. Validiti never sees the question, the records, or the answer.
Already have a draft from an LLM? Paste it in. Provenance labels every claim — VERIFIED, PARTIAL, NO SOURCE — against the same records. The label tracks the entity, not the prose: a made-up name in plausible writing still labels NO SOURCE.
Each query searches records covering hundreds of thousands of documents in roughly 200 ms. A full draft of a few paragraphs verifies end-to-end in single-digit seconds. Your LLM round-trip is on top of that.
Sealed binary, machine-bound, every internal package verified end-to-end. Tampering with the install refuses to start. Same Titus runtime defense that protects every Validiti SKU.
The live demo is a hosted instance running Provenance against three open medical record sets (FDA adverse-event reports, drugs, conditions). Both flows are live — ask a question, or paste an LLM draft to verify.
Try the verify flow first: a medical paragraph that mentions Metformin (real), Aspirin (real), and Validitomab (a made-up drug that doesn't exist) returns three labels in under three seconds — VERIFIED, VERIFIED, NO SOURCE. The made-up name has no escape route. Then try the ask flow with the same record sets and watch the grounded answer come back with citations attached.
Open the live demo →Two ask-mode scenarios and one verify-mode scenario. All three reproducible in the live demo.
14 adverse-event records returned, all with metformin + acidosis co-occurrence in patients over 65 with eGFR < 30. Citations attached to each.
3 review records flagged — none of them contradict the association; one notes baseline lactate as the differentiating factor.
Composes the final answer from those citations. Your LLM's context window holds all 17 records plus your prompt — it works only from what's there.
Metformin + lactic acidosis — strong support in adverse-event records.
Aspirin + gastrointestinal bleeding — well-attested across the record set.
Validitomab isn't in any medical record set — the LLM made it up. Caught before the draft ships.
Katz v. United States, 389 U.S. 347 (1967) — reasonable expectation of privacy doctrine. Verbatim opinion attached.
Smith v. Maryland, 442 U.S. 735 (1979) — third-party doctrine. Verbatim opinion attached.
Composes the answer from the cases that exist in your set. No fabricated citations possible — if a case isn't in the records, it doesn't reach the LLM. Your brief never quotes a case Provenance didn't pull.
Real numbers from the running engine. Reproduced on a 4-vCPU box, no GPU.
Latency is measured from text-in to labels-out, including reading the records the customer registered. Your numbers will depend on how big your record sets are and what hardware you run on, but the per-claim cost stays in the same range.
Three things people try when they want to catch hallucinations. Here's what each one actually delivers.
| Approach | What it actually does | Where it falls apart |
|---|---|---|
| "Use a second LLM to fact-check" | A second model reads the first model's output and gives a confidence score. | Both models hallucinate. The "checker" can be just as wrong as the original. No verifiable record set. |
| RAG with citation insertion | The LLM retrieves passages from a vector store and weaves them into the answer. | The retrieval may miss; the LLM may still mis-attribute or invent claims that look like they came from the retrieved passage. |
| Manual review by a domain expert | A human reads every line and approves it. | Slow, expensive, doesn't scale. Most drafts go out without it. |
| Validiti Provenance | Splits the LLM output into claims. Labels each one against records you control. Returns the original text annotated. Same engine answers a fresh question from records and hands a grounded prompt to your LLM. | Not a model. Not a vector store. Not a guess. The label is the truth or the record set's silence on it — your call. |
| Approach | Cost to verify 1M claims | Pricing model |
|---|---|---|
| Per-token LLM fact-check (GPT-4 / Claude) | $50 – $500 | $3 – $15 / M tokens · grows with every call |
| Vector DB + per-call grounding | $200 – $1,500 | DB hosting + per-query inference fees |
| Specialized eval / guardrails SaaS | $5,000 – $25,000 / mo | per-seat enterprise contracts, no per-call cap |
| Validiti Provenance · Personal | $19 / mo flat | first 90 days free · no per-token, no per-query |
| Validiti Provenance · Enterprise | $1,000 + $100 | $1,000/mo floor + $0.0001 per claim · published meter |
We're not saying don't use RAG or human review. We're saying you should know which claims actually have support and which don't, before either of those steps — without paying per token to find out.
Pick a tier. Your LLM stays yours. No per-token gotchas. Foundation is free for verified educational institutes; Personal is free for the first 90 days.
All tiers ship the same engine — same speed, same labeling logic, same sealed deployment. Higher tiers add team features for organizations running multiple machines.
A medical school teaching evidence-based research. A law school clinic running grounded case-law search. An undergrad program where every student needs to verify their own work.
Education tier · contactA medical writer on a deadline. An attorney drafting briefs. Three months to fold it into your workflow before any bill arrives.
Launching soonA regulated-industry editorial team — pharma, legal, compliance, technical writing. Every member runs Provenance against the same canonical record sets.
Launching soonA regional pharma editorial group. A law firm with multiple practice groups. Compliance audit-ready out of the box.
Launching soonMulti-region pharma, large law firms, regulated agencies. Self-serve published meter, no sales call. Verify a million claims a month for $100 over the floor.
Launching soonA legal research platform offering Provenance as a native verification feature. A pharma document system embedding it for every author. A regulatory compliance suite shipping it to every customer seat.
Launching soonEvery tier ships sealed. Every internal package verified end-to-end. Same Titus runtime defense as every other Validiti SKU. The same engine that runs the live demo runs in your install. No per-token fees. No per-query fees. No vendor lock-in.
The reason Provenance can be trusted is that the answer doesn't come from a model — it comes from records you control. We don't see your records, your text, or the labels.
