Products · WorldModel
A war-game simulator for the questions management cannot ask today.
VENETA WorldModel sits over a company's own operating systems in a read-only sandbox. It reads, simulates, gates and recommends. A person decides.
Overview
Simulate before you act.
WorldModel never writes to the network. It ranks causes from alarms and topology, simulates each action against the pessimistic edge of the service-level band, and drafts the recommendation with its evidence. Approval creates a change request for the operator's own system and an outcome memory. Nothing else changes.
Facts arrive only through tools and memory. The first step of every operational turn reads the live systems; recalled facts are stamped with when they were read and say to re-read. A critic, bound by the checks, fails closed.
How a question is answered
One loop, every time.
Read
Alarms, topology, readings and history through tools, inside the sandbox.
Simulate
Each candidate action against the key-performance band it would move.
Gate
Actions that fail the service-level gate are not indicated, and say why.
Approve and remember
A person approves. The change request goes to the operator's system; the outcome is remembered.
Architecture
Backup images in. One change-request draft out.
WorldModel reads your systems from point-in-time backup images that the VENETA Data Platform mounts read-only in an isolated sandbox. The ontology gives every tool and every memory one vocabulary, the agents ask the simulators, the memory layer learns under your approval, and the one thing that goes back to your systems is a change-request draft after a person approves. Nothing is written to the network.
Swipe sideways to see the whole diagram.
Read it bottom-up. Your business and operations systems are backed up as point-in-time images. The Data Platform mounts them read-only. WorldModel reads them through the ontology, the agents ask the classical, HPC and IBM Quantum simulators (the last through VENETA Sentinel), and your own LLM narrates and drafts. Your change management receives one draft after a person approves. Where a data platform such as Palantir Foundry already holds the integrated data, the integration layer stays and the decision layer sits on top.
Editions
One engine, an edition for each industry.
Editions share the engine. The domain module carries the tools, the seeds and the gates of that industry: telecom, energy, semiconductor, retail, manufacturing, defense.
What an answer carries
Evidence, not assertions.
- Readings
- Which systems were read, through which tools, and when.
- Simulation
- The band each option would move, with the pessimistic edge the gate was applied to.
- Gate verdict
- Passed or not indicated, with the service-level reason.
- Checks
- Grounding, length, simulation and quantum labels, tool use, and the critic's verdict.
- Quantum, when used
- Backend, qubits, shots, circuit depth and gap to the classical optimum.
- The model
- Which language model answered. The engine runs with the customer's own model, on premises.
How we measure it
Four gates, ten axes, no composite score.
veneta-worldmodel-bench is our proposal for measuring a WorldModel as a whole. Four gates come first: no data leaves the network, an authority level bounds what the system may touch, audit and roles, and reproducibility with honest labels. Then ten axes: accuracy, stability, security, forecast accuracy, simulation, optimization, self-evolution, self-improvement, memory and performance. Code grades every score, every number carries its conditions, and there is no composite score.
The proposal and the current card for veneta WorldModel, on project.veneta.ai
Enterprise-ready
The same engine, for the models you already run.
WorldModel's memory and self-improvement layer is also offered on its own, in front of the LLMs, agents and chatbots a company already runs, through an OpenAI-compatible interface. A lesson a model proposes is measured before it stays, approved by a person, and written to a ledger that an auditor can replay. Rejected lessons stay rejected.
The approval console, ledger, roles and dual control, learning budget, and undo are on by default. On top come the enterprise features: multi-tenancy (memory and policy separated per department, subsidiary or client), role-based access control, and single sign-on (SAML, OIDC). project veneta, the open-source layer, does not include those three.
Comparison
RAG, LLM wiki, and veneta.
RAG remembers documents, an LLM wiki remembers organized knowledge, and veneta remembers experience. They do not overlap, so you can use them together. veneta leaves facts where they live (tools and documents first), turns on a learned rule only after a person approves it, and records every change in a ledger you can rewind.
The full comparison and the measurements, on project.veneta.ai
Deployment
Inside your network.
- Editions
- The telecom edition runs the full loop today and the retail edition is in delivery to its first customer; other editions run on the same engine with their own domain module.
- Deployment
- Inside the customer's network, on the customer's hardware, in a read-only sandbox over existing systems.
- Data sources
- Today, point-in-time backup images mounted read-only by the VENETA Data Platform. The data platforms you already run, Palantir Foundry included, through connectors (planned).
- Language model
- The customer's own model, served locally. No data leaves to train anything.
- Quantum
- IBM systems through Qiskit, delivered by VENETA Korea Inc., which signed the IBM Quantum GSI (Global Systems Integrator) agreement.
Talk to us about your operating data.
Tell us which system holds the question you cannot ask today. A person answers every message.