Optimum v1

Optimum v1 is the industrial neural model developed for Cranc. Its purpose is to read machine behavior over time, recognize deterioration before it becomes obvious, support operator decisions with short reasoning, and work with deterministic safety logic rather than replacing it. It was built for factories where useful prediction requires more than a single sensor, a single machine type, or a single dashboard threshold.

Technically, Optimum v1 is a compact multimodal Decision Transformer-style industrial network with roughly 50 million parameters in its deployed decision core. The architecture follows the transformer idea of reading sequences rather than isolated rows: telemetry tokens, machine-state tokens, recent action context, return-to-go style operating targets, time-step embeddings, and optional text/image/audio-derived context are fused into a shared sequence representation. On top of that transformer backbone, Cranc uses specialized heads for action recommendation, outcome and downtime risk, health classification, reasoning class prediction, and compact operator-facing explanation signals. This makes Optimum v1 closer to an industrial sequence decision model than a conventional threshold model or single-sensor classifier.

The model was trained and refined through a staged industrial learning program. The corpus included dense machine time series, synthetic and real-world-inspired CNC sequences, minute-level machine histories, general mechanical fault records, electrical power patterns, coolant and pressure-flow cases, thermal and bearing scenarios, vibration and acoustic signatures, quality outcomes, operator approval traces, maintenance lifecycle records, and reasoning fields that describe why a decision should be made. The training program was designed around extremely large timestamp exposure, with datasets and preparation flows built for million-to-billion scale industrial time-series records as the corpus grows.

A key design choice is that Optimum v1 does not treat each row as an isolated fact. Industrial failure is often a shape, not a point. A machine can be normal at one second and already beginning to drift when viewed across a longer window. Optimum looks for relationships: whether current is rising faster than expected, whether vibration is changing relative to load, whether thermal behavior is consistent with the process phase, whether acoustic signatures match the learned baseline, whether quality signals are starting to degrade, and whether downtime probability is increasing even before the final alarm appears.

The model learned across many signal families: electrical power, voltage drop, phase imbalance, active and reactive power, motor current, mechanical load, torque, cutting force, hydraulic and coolant pressure, flow rate, bearing and spindle temperature, thermal drift, vibration RMS, crest-like behavior, acoustic emission, production cycle time, scrap rate, surface roughness, dimensional error, alarm state, operator action, maintenance timing, and machine lifecycle context. This breadth is important because the system is not meant to be a CNC-only classifier. CNC is one of the strongest proving grounds, but the same architecture is intended to generalize across industrial machines that produce time-series evidence of stress, wear, drift, overload, or abnormal operation.

Optimum v1 also learned from reasoning labels. The goal is not to make the model speak like a large language model. The goal is to make it give a compact operational explanation: vibration is rising faster than the learned envelope, coolant flow appears inconsistent with temperature, a tool-wear pattern is forming, a machine-specific baseline suggests the value is acceptable, or a shutdown risk should be reviewed. In product usage, this reasoning is translated by Cranc into operator-facing language so the decision is understandable without exposing internal labels.

The model works with Decadapter, the deterministic interpretation and interlock layer. This is one of the most important parts of the system. A neural model can detect patterns that fixed rules miss, but a fixed safety layer can catch obvious physical boundaries that a model may treat uncertainly. In internal evaluation, this hybrid design proved especially important on unfamiliar vibration data: the neural component could be conservative or over-alerted, while the deterministic layer prevented unnecessary shutdowns until a real physical shock threshold appeared. This separation makes the system more practical for production use because it does not depend on neural confidence alone.

Internal testing covered long-running simulated factory shifts, staged CNC deterioration, tool wear and breakage scenarios, electrical and power quality risk, coolant and pressure-flow failures, thermal and bearing issues, vibration-acoustic anomalies, general machinery faults, operator approval workflows, and out-of-distribution vibration files. In stress simulations, Optimum repeatedly detected deterioration before the configured failure point, sometimes minutes before a simulated stop event. Optimized inference tests also showed that the model can respond quickly enough for advisory industrial monitoring, while lighter deployments can still support useful review workflows when ultra-low latency is not required.

Optimum v1 is not presented as a magic system that can guarantee every future failure in every factory. The correct industrial claim is more disciplined: it can learn operating envelopes, detect changes across multiple signal families, surface early warning evidence, assist operators with decisions, and improve the speed and clarity of maintenance review. The value is strongest when a factory can provide enough historical and live data for baseline learning, and when rollout follows a responsible path: shadow mode first, advisory mode next, and deeper integration only after validation against the customer's machines.

The model was also designed to work with memory and active learning under control. A factory may have a machine where a globally unusual value is normal for that asset. Cranc can store tenant-specific context so recommendations are not based only on a generic threshold. Active learning can be used to adapt to a machine's normal behavior, but this must be gated: updates should be isolated per tenant or machine, tested against validation checks, and promoted only when they improve behavior without damaging existing capability. This is why Cranc separates observation, memory, model update, validation, and release.

The long-term direction of Optimum is multimodal industrial intelligence. Telemetry is the foundation, but images, 3D digital twins, audio, operator notes, quality reports, and maintenance forms all describe the same physical process from different angles. Optimum v1 establishes the core decision layer: read the machine, recognize risk, explain the reason, and work with a safety interpreter. Future versions can expand the same architecture into richer visual inspection, deeper sound analysis, stronger machine-specific adaptation, and more precise downtime and outcome forecasting.

The reason Optimum v1 matters is not only that it predicts. Many systems can raise alarms. Optimum's value is that it connects prediction with context: the machine, the signal family, the trend, the likely fault class, the recommended response, and the safety layer that decides how strongly the operator should act. That combination is what makes Cranc different from a conventional dashboard. It is not just watching values move. It is building an operational understanding of the machine over time.