Private · local · evidence-governed
For teams that can't accept a plausible but uncited answer. Every answer comes with its source, its time bounds, and its proof — or Weryta says it doesn't know. Runs on your hardware; your documents stay yours.
What an answer looks like
A question about a procedure at a point in time. Weryta returns a verdict, the exact document section, the interval it was valid for, its status, and tamper-evident identifiers — or an honest Unknown. (Example uses illustrative data.)
Procedure-X was approved.
Procedure-X is approved[wk 5, wk 42)VerifiedDirect evidence · proof bundle sha256:…
No — the record shows the opposite at week 45.
Procedure-X is-not approved[wk 42, ∞)VerifiedEvidence of falsehood — not absence of evidence.
Five verdicts, never a bluff
Most systems have two states: an answer, or nothing. Weryta has five — because "the evidence conflicts" is not "there is no evidence", and an association is not yet a claim.
The record directly supports it — with the source and the dates it held.
The record directly says the opposite — a withdrawal, a supersession.
The corpus holds the claim and its negation at equal specificity. Not missing knowledge — a conflict in the documents that a human has to settle.
Nothing is claimed. The associative layer links the concepts and points at where to look — a retrieval suggestion, never an answer.
No basis in the documents you gave it. It refuses to invent one.
"Derived" is not a verdict — it is a kind of answer inside a proof pack: an inference from parent facts, shown with its full chain and rule, never as a stored fact.
Why believe us
An external benchmark — bAbI (Facebook AI) — on questions the system has never seen, with no training at all. Weryta is not a language model: it has no pretrained knowledge, so any sentence outside its grammar is refused rather than guessed.
test questions (3 tasks × 1,000)
correct on every one of the tasks
claims asserted without evidence
Tasks: single supporting fact, time reasoning, deduction by inheritance. For scale, an untrained LSTM scores roughly 50/27/21% here. Separately, on an internal range of worlds generated from hidden semantics that candidate rules never see, one gate is absolute: zero false assertions — a rule that states an untruth even once does not pass, no matter how much it gains elsewhere.
Weryta would rather say Unknown than guess — and we have a number for that, not a slogan.
Decision ledger
Every decision to accept or reject a fact is appended to a ledger in which each row seals the previous one with a cryptographic hash. Cut a single row, or change a single character, and the chain no longer adds up — visibly, immediately. A mistake is never fixed by overwriting: it is a new row.
The head of that chain is anchored into Bitcoin via OpenTimestamps. We store no data there — only the hash is anchored. Thirty-two bytes leave the building. Your documents never do.
Rewriting our history would mean forging a Bitcoin block. Verification runs offline, against your own node — you do not have to trust us to check it.
A second witness
A fact about when a rule came into force is accepted automatically only when an independent state registry confirms it (the Polish parliament's ELI interface). A model can be confidently wrong; a registry is a witness that does not guess.
From our own ledger: the model gave an effective date of 2020-01-02,
the registry answered 2020-01-06. The fact never entered the store.
Two models agreeing is not proof. Agreeing with the state registry is.
Two axes of time
Weryta separates valid time (when a rule was actually in force) from knowledge time (when we learned of it). That is what lets it answer the question audits actually ask: "what was true on 1 March 2021, as far as anyone knew on 1 June 2021" — rather than only "what is true today".
Two axes of time are decades-old computer science (bitemporality, SQL:2011) and we did not invent them. What Weryta adds is a proof pack you can verify offline, without trusting us.
What you need to run it
Weryta is a small Rust binary that runs offline on hardware you already own. Checking a proof pack needs no graphics card and no network either.
Heavier hardware is required only if you want to run local language models in the learning pipeline — Weryta itself, and proof verification, do not.
Read both columns
How it learns
The system learns by growing the set of verified facts — not by tuning weights on the path to an answer. The associative layer helps find the right documents; what gets asserted is decided by evidence alone.
The channel through which frequent use could have quietly flipped a verdict was measured, judged dangerous, and closed at the compiler level — by types, not by policy: the field is private and read-only, so code that could rewrite truth through usage simply does not compile.
What we are working on
Who it's for
Teams that live in technical procedures and version changes, administrative decisions, registers and deadlines, quality documentation, incident reports, maintenance procedures, or internal audits — where the real question is not "what's the answer" but "show me the document, the section, and the version that was in force that day."
Collaboration
The collaboration is about method and verification — it does not change the rule that only verified evidence changes what the system believes.
When did someone last answer a question about a procedure, decision, or condition — but couldn't show the document, the section, and the version in force on that day?
Book a 25-min slot →Prefer email? contact@weryta.com — or just copy the address. We reply within a day.