Predictive and preventive systems are forecasts with stated uncertainty for grid and asset operators, so that they act before a peak or a failure and can check the forecast afterward.
Fix the prediction question and reconstruct only eligible evidence.
Dated exposure/condition history under an explicit cutoff.
Named domain target, units, cutoff and horizon.
Eligible later evidence for a subsequent estimate, preserving the earlier forecast.
Fix the prediction question and reconstruct only eligible evidence.
Population, event/target, units, cutoff and horizon.
Exposure, incomplete observation and event history.
Only evidence eligible at the declared cutoff supports the prediction.
Separate operating-load response, persistent condition difference and unreliable sensing.
Eligible channels and incompleteness.
Separate operating-load response, persistent condition difference and unreliable sensing.
Comparable load and source timing.
Competing explanations and sensing limits.
A persistent difference and unreliable sensing call for different next steps.
Apply the appropriate explicit domain basis without merging unlike targets.
Eligible domain evidence and target definition.
Load/sensing interpretation and remaining uncertainty.
Explicit class/local or task-specific basis, never a universal shared target.
Apply the appropriate explicit domain basis without merging unlike targets.
Asset survival/class/local evidence and maturity.
Separate declared episode target and resource use.
Asset condition and task outcome use distinct targets and domain evidence.
Carry the partition, uncertainty and support conditions with the estimate.
Estimate, horizon and evidence maturity.
Declared calibration/support conditions.
Carry the partition, uncertainty and support conditions with the estimate.
Grouped partition, prior strength and nominal interval.
Applicable population and unresolved coverage assumptions.
Uncertainty retains its calibration and population-support assumptions.
Compare informative observation, inspection or intervention proposals under real constraints.
Immutable prediction record.
The distinction a next observation could resolve.
Feasible options and consequences; no action authority.
Compare informative observation, inspection or intervention proposals under real constraints.
Domain-specific options and unresolved question.
Relevant criticality and operating consequences.
A proposed intervention remains separate from equipment authority.
Freeze the prediction before the decision and compare genuinely later evidence.
Estimate with uncertainty and explicit limitations.
Selected response appended beside the earlier forecast.
Independent later evidence under the original prediction identity.
Actual outcome for the declared target when available.
Freeze the prediction before the decision and compare genuinely later evidence.
Original target, horizon, estimate and uncertainty.
Later observation and separately adjudicated outcome.
The forecast stays frozen while later observations and adjudicated outcomes are added.
Dated exposure/condition history under an explicit cutoff.
Named domain target, units, cutoff and horizon.
Eligible later evidence for a subsequent estimate, preserving the earlier forecast.
Fix the prediction question and reconstruct only eligible evidence.
Population, event/target, units, cutoff and horizon.
Exposure, incomplete observation and event history.
Only evidence eligible at the declared cutoff supports the prediction.
Eligible channels and incompleteness.
Separate operating-load response, persistent condition difference and unreliable sensing.
Comparable load and source timing.
Competing explanations and sensing limits.
A persistent difference and unreliable sensing call for different next steps.
Eligible domain evidence and target definition.
Load/sensing interpretation and remaining uncertainty.
Explicit class/local or task-specific basis, never a universal shared target.
Apply the appropriate explicit domain basis without merging unlike targets.
Asset survival/class/local evidence and maturity.
Separate declared episode target and resource use.
Asset condition and task outcome use distinct targets and domain evidence.
Estimate, horizon and evidence maturity.
Declared calibration/support conditions.
Carry the partition, uncertainty and support conditions with the estimate.
Grouped partition, prior strength and nominal interval.
Applicable population and unresolved coverage assumptions.
Uncertainty retains its calibration and population-support assumptions.
Immutable prediction record.
The distinction a next observation could resolve.
Feasible options and consequences; no action authority.
Compare informative observation, inspection or intervention proposals under real constraints.
Domain-specific options and unresolved question.
Relevant criticality and operating consequences.
A proposed intervention remains separate from equipment authority.
Estimate with uncertainty and explicit limitations.
Selected response appended beside the earlier forecast.
Independent later evidence under the original prediction identity.
Actual outcome for the declared target when available.
Freeze the prediction before the decision and compare genuinely later evidence.
Original target, horizon, estimate and uncertainty.
Later observation and separately adjudicated outcome.
The forecast stays frozen while later observations and adjudicated outcomes are added.
An illustrative method for retaining a forecast before comparing independently supplied later evidence.
Additional mechanisms are described in the technical reference.
What the research says
Weather services show that learned forecasts can beat physics models on most targets while costing far less to run. Extrasensory applies the same discipline to equipment and load: each forecast states its uncertainty, is stored before the outcome and is scored afterward.
Two assets can have the same service interval and different operating histories. A signal can rise because load changed, because condition deteriorated or because the measurement is unreliable. Our predictive architecture keeps those explanations separate before relating them to an intervention.
Longitudinal histories retain operating exposure, event time, changing conditions and incomplete observations. Surviving assets contribute evidence without being labeled as failures.
Equipment-class priors provide a cold-start basis. Condition adjustments add local evidence while keeping the inherited assumptions visible.
The model retains horizon, applicability and uncertainty when comparing inspection, maintenance and continued operation.
Stored forecasts precede later outcomes, preserving the basis for calibration and intervention review.
Extrasensory’s general-purpose architecture carries explicit longitudinal training geometry. The maritime predictive/preventive application is one operating context; domain adapters retain their own equipment and observation rules.
A history ending without a failure is not a negative event at every future horizon. An asset first observed part-way through its life enters the risk population at that later time.
Different events may terminate a period of exposure, while a repaired or repeatedly serviced asset can produce further event intervals. Their labels and risk sets must retain that distinction.
Predictor inputs are separated from future outcome information. Grouped partitions keep evaluation populations and calibration assumptions explicit.
A calendar or hours-based countdown can coexist with a survival estimate. The architecture does not collapse either into the other.
The grid operator decides whether and how far to shift load; the forecast supplies the band. Illustrative.
Illustrative situations, connected to the mechanisms that make them possible.
Marine engineering
Load, condition and sensor histories change while equipment remains in service.
Industrial operators
An asset has limited local failure data but belongs to an understood equipment class.
Asset-intensive organizations
Inspection, continued operation and maintenance carry different practical constraints.
The system can combine class information with operating observations while showing which part of the estimate comes from each.
A cold start uses an equipment-class basis. Accumulated applicable evidence changes the local estimate and its stated maturity rather than inventing local failure labels.
Load normalization and condition channels distinguish an operating regime change from an anomaly. Bounded feedback tuning changes declared parameters within limits.
Finite-sample rank-based calibration retains its partition and statistical assumptions. The existence of a conformal method does not create unconditional coverage under arbitrary distribution shift.
The intervention layer compares feasible actions under operational constraints. A high predicted risk does not itself grant actuation authority.
Vagus predicts whether a defined episode ends well and how many resources it is likely to consume. Its proxy/0.2 target is an episode completed without an observed interruption or red-test/failure ending; independently labeled outcome_ok is separate. Neither is an automatic measure of task correctness. It does not apply an asset-failure model to a software task.
A Beta–Binomial hierarchy links portfolio, holder, repository, task kind and worker. Sparse cells inherit a bounded prior from related observations without counting the same branch twice.
Task features available before execution.
A declared outcome probability and evidence basis.
Log-normal models produce token and wall-time medians with 80% intervals in the inspected implementation. These are nominal model intervals; estimates remain distinct from budgets, observed coverage and guarantees.
The ledger scores a prediction made before its outcome. Unknown endings stay unlabeled; target definitions and a running-base-rate comparison remain explicit. Brier score, log loss and reliability bins are compared with a running base rate; uncertainty is clustered by root session and subgroup verdicts are Holm-adjusted.
Manna’s typed world-model boundary links a forecast to an objective, scope, horizon and expiry before a bounded proposal can be admitted.
A scoped belief and current grant.
A proposal whose motivating forecast can be compared later.
The research organ’s prediction is separate from action authority and installed adoption.
In an illustrative operating episode, a vessel’s condition signal rises during a higher-load period. The explanation follows the operating history, the applicable condition model and the uncertainty around the decision.
The intervention record distinguishes the load-normalized evidence from the maintenance countdown, names the forecast horizon and compares the available response. A later inspection can supply a new observation without rewriting what the system knew when it made the earlier recommendation.
That is the common technology path from prediction to prevention. Data sources, event definitions, calibration population and intervention permissions remain specific to the commissioned domain. The general-purpose edition does not inherit every vessel-specific implementation by naming it.
Inputs, retained state, outputs and limits of Naevum’s Predictive & preventive systems architecture.
Evidence of calibration and usefulness for the specified domain, horizon and decision. Prospective evaluation should include false alarms, missed events and whether the response was useful.
The general predictive service and its maritime application retain explicit capability mappings. Longitudinal labels, leakage policy and calibration population are part of the model definition. Federated and representation-learning research is not an operating capability claim.
Vagus’s episode target may be a completion proxy. Its prequential replay respects time order but does not by itself establish prospective downstream utility or a causal worker ranking.
Nominal resource intervals, conformal coverage and survival uncertainty refer to different models and populations. Calibration evidence is specific to the target, horizon and observation process.
For a commissioned domain, the technical contract records data eligibility, event definition, censoring, applicability, prediction record and available interventions. No accuracy, downtime or financial improvement figure is inferred from implementation alone.
Which peak or failure would you act on earlier, with a forecast you could check afterward?
Explore predictive operations