Repairing world models was only the beginning.

01 · The starting point

Models become wrong.

A model is trained under a particular operating regime. After deployment, equipment wears, materials change, actuators lose authority and relationships between commands and outcomes evolve. Most of the system may remain correct while one local assumption is no longer true.

A08 started with a focused research question: when a deployed world model becomes wrong after a hidden change, can we repair it without retraining the entire model?

That research became Interactive World Model Repair, or IWMR. It detects a mismatch, prepares a local typed correction, verifies the candidate before activation and creates an explicit model version with rollback.

The objective is not to let an AI rewrite itself freely. It is to make adaptation observable, governed and reversible.

02 · What IWMR established

A repair mechanism became a working system.

Our principal confirmed result used an action-conditioned dynamics model in the official Gymnasium MuJoCo Reacher environment. A hidden change reduced the effect of one actuator: the command remained valid, but it no longer produced the motion predicted by the model.

The same byte-identical repair Core was later integrated with a different learned relational dynamics architecture containing 10.07 million parameters and a different family of typed corrections.

This established an important boundary: IWMR was more than a repair script written for one robot task. But it also revealed what was missing.

03 · The larger problem

A world model does not create value in isolation.

An industrial company does not begin with a production-ready world model waiting to be repaired. It begins with PLCs, sensors, actuators, SCADA systems, historian data, operating procedures, constraints and a business KPI that needs to improve.

Before a model can be repaired, it must be defined, connected, trained, validated and operated. The industrial question is therefore larger:

This is the transition A08 is making. IWMR remains the technical core for governed adaptation. Around it, we are building the path from industrial signals to measurable operational decisions.

04 · The industrial layer

Above SCADA. Before action.

A08 is designed to sit behind existing industrial systems, not replace them. The SCADA remains the operational system. The HMI remains the operator interface. A08 adds a predictive decision layer that can initially run entirely in shadow mode.

We do not believe an industrial world model needs to understand an entire factory from day one. A useful deployment can begin with one process, one set of controllable actions, one KPI and explicit operating constraints.

Once one model creates measurable value, the same data and operating layer can support additional KPIs, production lines and sites.

05 · First demonstrator

A complete loop on an automotive curing process.

Our first demonstrator is a simulated automotive paint-curing line: a three-zone thermal oven, a controllable conveyor, delayed temperature dynamics, energy consumption, production throughput and a quality outcome for every processed body.

The plant publishes public telemetry through OPC UA. An edge collector stores it in a historian before the model, planner or interface can use it. The model cannot access private simulator state.

Improving a healthy process

One qualified recommendation reduced energy per good body by 4.76% while maintaining 38 bodies per hour and 100% simulated quality. Another increased throughput from 40 to 42 bodies per hour while preserving quality, with its additional energy cost and €3.66 contribution break-even made explicit.

Remaining useful after the process changes

In another scenario, the conveyor continued operating but stopped responding to commands as expected. The old model predicted output the process could no longer deliver. Production fell to 28 bodies per hour.

After the changed relationship was identified through the governed workflow, the system proposed a revised command that recovered 36 bodies per hour while preserving the simulated quality constraint.

06 · What comes next

From a research mechanism to an industrial product.

The direction of A08 has expanded, but the underlying thesis has not changed. We started with model repair because adaptation is one of the hardest and least governed parts of deploying learned dynamics.

We are now building the surrounding system because repair only becomes commercially useful when the model is connected to a real operational objective, trustworthy temporal data, explicit constraints, human authority and measurable outcomes.

The next meaningful step is not a larger simulation. It is applying this architecture to one bounded industrial KPI with accessible operational data, an agreed baseline and recommendations first evaluated in shadow mode.

One process. One KPI. One model. Then expand.