The engine that keeps facts with their source and date.
Tyla stores facts with their source and the period in which they were valid. Rules a person can read derive the answers, and every answer shows the facts and rules behind it. It runs in your data centre or in the EU.
Three applications you can build with Tyla.
Which terms applied on a given date?
Every contract, amendment and price list with its validity period. You get the terms that applied to an order on its date, with the document that set them.
Which obligations are due, and since when?
"Report within 30 days of an incident" becomes a rule with a time window. Tyla lists every case that meets it on a given day.
Which partners are critical today?
Lists, registers and ownership data change. Tyla keeps every version and flags a supplier that was on a sanctions list in the last 90 days.
Data comes from your databases, files and documents. A language model can read documents. Where rules exist for a kind of fact, Tyla validates the extracted facts against them before it stores them.
Tyla builds every answer from stored facts.
A language model holds implicit knowledge: tendencies in its weights, without a source or a date. Tyla works with explicit knowledge: every fact with its source and the period in which it was true. In 2028 you can show an auditor exactly what you knew in 2026.
- Cited. Every fact links to the document or record it came from.
- Dated. "What did we know on 14 March?" is an ordinary query.
- Corrections keep their history. A corrected fact is stored as a new version with its date. The old version stays, so earlier decisions remain explainable.
- Not on any sanctions list
- On the EU sanctions list
- Removed from the list
Example with an invented supplier. A new version never overwrites the old one.
Time is part of the rules.
Rules in Tyla can refer to time. From the validity periods of the facts, Tyla calculates how long a conclusion holds. Here is one rule, applied to the supplier from above.
A supplier is critical if it was on a sanctions list at any time in the last 90 days.
The answer on 15 April: Nordtek is critical, until 29 May.
Nobody entered the date 29 May. Tyla calculated it from the last day on the list and the 90-day window, and gives both as the reason.
How Tyla compares.
A language model answers from its training data. Retrieval-augmented generation (RAG) adds documents found for the question. Graph-memory libraries such as Graphiti or Cognee attach timestamps to a graph database. Tyla stores time in the data model itself, so an answer can be replayed and cited for any date.
| Aspect | General language model (LLM) | RAG pipeline | Graphiti / Cognee | Tyla |
|---|---|---|---|---|
| How time is handled | Fixed at the training cut-off | Only what was retrieved for this question | Validity timestamps on edges | Bitemporal, part of the data model |
| Reasoning | Predicts the next word | Retrieval only | Graph lookup; entities extracted by a language model on import | Deductive, with the temporal logic DatalogMTL |
| Auditability | Practically none | References to documents | The graph can be browsed; derivations cannot be replayed | Every derivation can be replayed for any date |
| Deployment | Usually cloud only | Depends on the model | Self-hosted or in the vendor's cloud | In the EU, on premises or air-gapped |
The research behind Tyla.
The rule language is DatalogMTL: Datalog, a rule language from database research, extended with metric temporal logic (MTL). Tyla continues the Temporal Vadalog System, a DatalogMTL reasoner whose principal developer was Markus Nissl during his doctorate at TU Wien.
critical(S) :- ◇⁻[0,90d] sanctioned(S).- How intrusion detection systems handle timePaper, arXiv, 2026
- Preventing invented answers: measured resultsWhitepaper, 2026
- Preventing invented answers: the architectureWhitepaper, 2025
- Author profiles on DBLP: Markus Nissl, René Heinzl
What buyers ask before they start.
No. Tyla answers from stored facts and rules. A language model is optional: it can turn a question into a query or read documents, and you choose which one.
Your experts decide what a rule says, and in the pilot we write the first rules with them. The rules are readable, so your team can check and change them later.
One question from your work and a sample of the data it needs. We write the first rules with your team and show you the answers with their evidence.
Per year, for your installation. After the pilot you receive an offer based on what the pilot covered.
Start with a pilot on your data.
We load a sample of your data into Tyla, write the first rules with your team and show you the answers with their evidence. After the pilot you decide: an annual licence, or a project with a fixed scope.
- Runs
- In your data centre or with EU hosting
- Your data
- Not used to train models
- Language model
- Your choice
- Built in
- Vienna, Austria