The Zendrix world model evolves: verifiable intelligence for critical decisions

The Zendrix world model has just evolved.

Our technology can now follow world state with persistent memory, anticipate scenarios with calibrated uncertainty and record — for audit — why each decision was made. In practice: more perception, more context and more reliable decisions in critical environments.

It is another advance in building Zendrix's sovereign intelligence.

Systems that sustain critical decisions share a silent requirement: getting the next forecast right is not enough. You must know what you know, record why you know it, and declare — before acting — what you do not know. A purely predictive model, however accurate on average, offers none of these guarantees. In domains where decisions carry physical consequence, that difference stops being academic and becomes operational.

The Camile World Model (CWM) was built for exactly this class of problem. It maintains a persistent representation of world state — observations, beliefs with history and explicit confidence, relations and forecasts — and treats every new piece of information as something that may confirm, revise, or refute what was previously established. Reading never changes the world; every result is reconstructible and auditable. In 2026, the research program behind the CWM completed its most rigorous cycle: version 1.0.0 declared all success criteria measured, re-executed its full verification chain, and publicly documented its limits — with the path to surpass them.

What sets a world model apart from a predictor

The difference starts with the unit of decision. Instead of answering "what comes next?", the CWM answers "what state is the world in, what changed, and what is most likely — with what uncertainty band". Every forecast is born with a calibrated interval and is verified when the future arrives: pending forecasts are confronted with the real event tape, and the record of hits and misses forms an auditable ledger. The system also recognizes when context shifts — calm, turbulent, or transitional regime — and adapts its reading without breaking the continuity of memory.

This discipline produced measurable results in public evaluations: state persistence with full retention in structured-memory benchmarks, world tracking accuracy above 0.94 under an independent protocol, and causal reasoning evaluated on the association–intervention–counterfactual ladder under an official judge. Where performance was not superior, that is published too: evidence is what separates a promise from a capability.

A decade of gaps, closed family by family

The CWM research program organized its known limitations into an explicit inventory: each gap received a family, a target, and a formal criterion for being surpassed — a statistical method defined before results. Across successive versions, five families were attacked and closed: seasonal worlds, long-memory worlds, oscillatory worlds, non-stationary worlds, and composition. The last of them ended the phase with the inventory walked through completely: the map of what the model could not do is now fully measured.

Closing a gap here does not mean improving an average. It means a class of behavior that previously resisted now holds under the pre-registered formal criterion — verifiable by any third party with the same methodology. And where the promotion process demanded extra rigor, it was unified: every new capability crosses the same statistical gate before entering service.

Efficiency as a mission parameter

World intelligence in critical environments lives with constrained resources: energy, hardware, intermittent or absent connectivity. The CWM was measured on this axis too. The public watt-hour efficiency curve covers six operating points — from the minimal edge profile to the full one — each with its reproducibility signature. The complete organism, including world, beliefs, and accumulated learning, operates in 256 KiB: it survives cold restarts, rebuilds its state from the durable record, and never relearns what it already knew. For operators, this turns "how much intelligence fits my envelope" into an explicit choice, with numbers.

What this changes for those who operate critical systems

For an institution that must decide under uncertainty, the properties translate into concrete capabilities. Planning stops being a black box: the system exposes the state behind each recommendation, forecasts with their bands, and detected contradictions. Proposed actions can be simulated in alternative worlds before the real decision — counterfactual scenarios on the same operational data. And the audit record follows the entire cycle: what was observed, what was believed, what was forecast, and what actually happened.

None of it depends on blind trust in a vendor. Benchmarks are public, with digital signatures and declared reproduction; the service contract is compatible by principle — extensions are additive and existing integrations do not break. The architecture was designed for inspection: every result carries the engine version that produced it, and legacy behavior remains reproducible.

Limits, declared

The CWM does not promise omniscience. World models work with what can be observed: statistical association is not causation, and the system itself records when a question would require a laboratory the data cannot provide. Published results are specific to the datasets and windows tested. And the gap map — though complete — is an inventory of what has been measured, not a certificate of completeness about the unknown.

Trust, here, takes the shape it should have in engineering: not the absence of declared failures, but the presence of mechanisms to detect, measure, and correct them. In critical decisions, that is the only kind of trust that holds.