Technology · Architecture
From governed evidence to scored outcomes.
Your systems stay in place and in command. Kalyr keeps the reference, provenance and permission behind every use, with predictions and observations as separate records.
The shape of it
Three planes over one moving world.
Your process keeps running. Its evidence is referenced above it, joined to where it lives. The planned learned layer sits above both, and results come back to be scored.
Your systems stay the source of truth and supply most of the state. People add what their systems miss, when the decision needs it.
The target loop
Eleven steps. One loop.
Each step is recorded separately, so every error traces to the step that made it. Steps marked planned are in development.
-
Telemetry, records, documents, video, human reports and external data, each where it already lives.
Source · time · context -
What exists, where it lives, its source and its time. The reference travels; the evidence stays home. Access stays with permission.
Source · location · time -
Each use is checked against purpose, scope and expiry before a record is read. A withheld source is held at your boundary. Zero reads.
Purpose · scope · expiry -
Kalyr reads what the permission allows, and records the read itself.
Permitted read -
A thin mapping reads each source where it lives, so local systems stay in place. Machine evidence stays attributed to the machine.
Entities · states · actions · outcomes -
The situation at the decision point, each element marked observed, reported or inferred, with its uncertainty. Your systems stay the source of truth.
Belief · uncertainty -
For each candidate move, the predicted next state and its range.
Candidate options -
Sealed before anyone acts and kept exactly as it was. Your team decides.
Seal time -
The next state arrives as its own record. Options not taken stay predictions.
Fixed horizon -
The chosen prediction is scored against what happened, then valued under the contract your finance team owns.
Prediction score · value contract -
Your team reviews each episode before it may teach the next prediction.
Review · admission
Open any step for the detail
What is built
One episode contract. A pack per decision.
One episode contract runs under every loop: keep your integration when the model or site changes. A decision-family pack carries what one decision needs, so a proven deployment repeats on the next line. You own the value contract, so you can audit every result.
Built · the same across models and sites
Built per decision family
The pack
Each pack is written down. Evidence can show where it is wrong.
When evidence repeatedly separates things the pack treats as the same, a versioned change can be proposed. People decide.
Disagreement.
The evidence and the pack diverge.
Proposed revision.
A specific change, with the evidence that produced it.
Reviewed version.
People decide. Earlier facts keep the version they were written under.
Control
Two questions before every read.
What exists, and where does it live? That is the evidence reference. May it be used for this purpose, now? That is permission. They are answered separately, for every person, system or agent.
Every record keeps its source, its permission and its purpose.
Recherche
The same episode contract may carry richer state across teams and environments, where permission allows.
Suivant
Compound your intelligence.
See one decision
Follow one decision from state to sealed prediction to settled result.
Get started
Two minutes. Start from what you already know.
Lausanne, Suisse