The contradiction that costs you the audit is already in your files.will find it.
Hosted in Germany
The stack runs on German infrastructure, and every customer runs on their own.
Zero training promise
Your documents never train a model. Not ours, because we train none, and not a provider's.
GDPR compliant
Processing agreement and the list of subprocessors on request, before deployment.
EU AI Act
A decision support tool for qualified professionals, and documented as one.
Hosted in Germany
The stack runs on German infrastructure, and every customer runs on their own.
Zero training promise
Your documents never train a model. Not ours, because we train none, and not a provider's.
GDPR compliant
Processing agreement and the list of subprocessors on request, before deployment.
EU AI Act
A decision support tool for qualified professionals, and documented as one.
What is KNTRA?
KNTRA is software that checks documents against each other and returns the passages that contradict one another.
A supervisory rule against a bank's own policy, a protocol against the production record, a master agreement against the addendum that was signed two years later.
What comes back is a list of findings rather than a summary or a score.
Why KNTRA?
Nobody holds a rulebook in their head. Every sentence carries its obvious meaning, a quiet implication and a cross-reference, and there are dozens of them on every page. Check a hundred-page policy against a hundred-page regulation and you are looking at more than a million pairs that all have to keep agreeing, every time either document changes.
So in practice, review means sampling. KNTRA reads every pair, using formal logic where it decides instead of probabilistic guessing.
Built for
Finds nine in ten.
The German regulator published its MaRisk revision as a change version with 529 marked changes, 240 of which carry a real contradiction. KNTRA returned 212 of them. That is 88 percent, measured on public material, with the misses in the report.
Your human in the loop
- Months become minutes
- A check does not end in a report you file away. It opens a working view of every finding, each one a click away from both source passages.
The machine does the months of searching, and one person spends a minute per finding, judging instead of hunting. KNTRA hands over what it found in two categories, and your expert has the last word.
-
Contradiction
Both quotes were found word for word in the files they name.
-
Needs review
Only one side could be located, so it goes to a person before it counts.
Your data stays yours.
- A stack of your own
- One customer is one running instance: its own container, its own database, its own address. Nothing is shared but the edge that routes the request, so there is no common store for a query to reach across.
- On German infrastructure
- The stack runs in Germany, on infrastructure we operate ourselves. The reading step currently uses a model provider outside the EU, configured for zero retention and zero training on every call, and we would rather you check that than take it on trust.
- Nothing trains a model
- No customer document is used to fine-tune or train any model. This one is structural as well as contractual: training on your documents would make us the provider of the model rather than the operator of a system, which is a position we have designed ourselves out of.
- Yours to keep or delete
- You set how long finished checks are kept, and deleting a run removes it and records the removal.
Frequent questions
What happens to our documents?
Each customer runs on their own stack, with their own container, database and address, so there is no shared store for a query to reach across. The reading step runs through a model provider outside the EU, configured for zero retention and zero training on every call; the processing agreement names every processor involved before deployment, and you set how long finished checks are kept.
Does KNTRA train on our data?
No. Nothing you upload is used to train or fine-tune any model, and that stands in the contract.
Which documents can KNTRA read?
PDF files that carry real text, and plain text files. A PDF holding only images is refused with a message naming the reason, because a scan has no text to quote.
Does KNTRA decide anything for us?
No, and the boundary is deliberate. KNTRA returns the contradiction and the two quotes behind it; whether to correct a clause, escalate it or accept it stays with the person who owns that decision.
Why not a team of people, or just an LLM?
People are excellent judges of a single contradiction and hopeless at holding a hundred pages against a hundred pages, because the pairs run into the millions and attention does not. A language model reads fluently, holds only what fits into one context window, and sounds equally confident when it is wrong. Four things separate a KNTRA check from both. The proof: every quote must survive a character-by-character lookup in its source file before you see it, and what fails the lookup is discarded and counted, never reworded. The coverage: a check reads all documents of a run together, and thirty files fit into no chat window. The process: every check follows the same recorded steps, is logged, and exports with its evidence attached. The commitment: we stand behind this with the pilot price, not with an adjective.