TECHNICAL NOTE · TECHNOLOGY
The integrated decision object
What changes when the model returns more than a prediction?
OCTOBER 2026
7 MIN READ
STANDARD DEEP MODEL
- INPUT
- MODEL
- PREDICTION
One output. Everything else is decided around it.
XPDEEP MODEL
INPUT
→MODEL
→CONFIGURED DECISION OBJECT · AVAILABLE PRIMITIVES
- PREDICTION / ESTIMATE
- STRUCTURAL EXPLANATION
- ATTRIBUTED DOMAIN STATE
- MINIMAL FEASIBLE INTERVENTIONwhere applicable
- STRUCTURAL JUSTIFICATION
The complete decision object requires the relevant capabilities to be supported, retained and configured; these are not five separate products.
Machines are starting to carry decisions with real consequences, adjusting a process, derating an asset, gating a perception output. Each of those decisions rests on what an AI model returns.
Most deep-learning systems return a prediction.
That may be a probability, a classification, a state estimate or a forecast.
For many applications, that is enough.
But when an AI model participates in a consequential machine decision, the surrounding system needs more than the output itself.
It needs to know:
- What drove this output?
- Is the model still operating inside conditions it learned?
- If the outcome needs to change, what feasible input change could alter it?
- Why does that intervention change the predicted outcome?
Xpdeep brings those questions together into one model-level decision object.
The important distinction is not simply that the model explains more.
It is that the output becomes structured enough to participate in a bounded operational decision.
01 THE OBJECT
The prediction plus four primitives. One decision object.
01 · STRUCTURAL EXPLANATION
What drove this output?
The learned model structure that produced the output: which input variables, with what weight, through which temporal dependencies.
Not reconstructed afterwards by a surrogate.
02 · ATTRIBUTED DOMAIN AWARENESS
Is the model still valid here?
A graded distance to the trained domain, with the responsible variables and temporal windows attributed. Includes novel-class detection where relevant.
Not a confidence score.
03 · CONSTRAINED MINIMAL INTERVENTION
What is a minimal feasible move?
A feasible input change, optimized for minimality, that improves the target outcome under stated operational constraints.
Not a recommendation. Not a global optimum.
04 · STRUCTURAL JUSTIFICATION
Why does this move change the outcome?
The learned structural dependencies that explain why the intervention changes the predicted outcome.
Not a narrative rationalization added afterwards.
The configured decision object brings together prediction, structural explanation and attributed domain state. A separate optional inference adds a minimal feasible intervention with structural justification where supported and retained. Native structural explainability is the foundation; the three loops productize it, with value in design and diagnosis as well as bounded authority.
The prediction is not traded away for this: on time-series data, accuracy is at least equivalent to standard deep learning and frequently superior.
The value is in the integration, not in any single primitive.
The primitives are not independent add-ons.
1
PREDICT + STRUCTURAL EXPLANATION
2
CHECK DOMAIN
3
WHEN AN INTERVENTION IS USEFUL: PRESCRIBE
4
JUSTIFY THE INTERVENTION
→
CUSTOMER-DEFINED RULES →ACT · HOLD · ESCALATE
Primary inference produces prediction, with structural explanation when explanation generation is retained.
Configured attributed domain state characterizes the input’s relation to learned conditions.
If an intervention is useful and the modality supports prescription, a separate inference computes a minimal feasible change under constraints.
The model then returns the structural justification of why that change alters the predicted outcome.
RUNTIME
Runtime explanations are configurable.
PRIMARY INFERENCE
SAME FORWARD PASS
- INPUT
- MODEL
- PREDICTION + EXPLANATION / DOMAIN STATE WHERE RETAINED
OPTIONAL PRESCRIPTIVE INFERENCE
SEPARATE INFERENCE · RUN ONLY WHEN USEFUL
- CURRENT STATE + CONSTRAINTS
- MINIMAL FEASIBLE INTERVENTION + STRUCTURAL JUSTIFICATION
When explanations are retained, prediction and structural explanation share one forward pass; configured domain state is read from latent space, not a second model. A production model may omit explanation-generation functionality because the application does not require runtime explanations or to reduce its resource footprint. Development-time native structural explainability remains foundational, but the complete decision object is then unavailable. Domain awareness, prescription and structural justification depend on the capabilities actually supported and retained. Prescription always runs as a separate optional inference.
NATIVE · COST AT INFERENCE
- When explanations are retained, prediction and structural explanation share the same forward pass. The explanation adds no extra compute or latency. The explanation layer’s memory footprint must be budgeted.
RETROFIT · COST AT INFERENCE
- The resulting artefact is larger than the original; compute and memory overhead are measured on the target.
FIELDS AND TYPES
The exact structure of the decision object, its fields and its types are specified in the developer documentation. Inference APIs are available in C / C++ and Python, with outputs exportable in JSON and Parquet.
Open developer documentation ↗The explanation comes from the learned structure itself.
VARIABLE 03
T−18 → T−06
LOCALIZED WINDOW
MORPHOLOGY
DEPENDENCY
On temporal data, Xpdeep does not stop at ranking variables.
It can localize the time windows and signal morphology that carried the model output.
This lets engineers inspect the learned decision at signal resolution.
POTENTIAL ENGINEERING USES
MODEL DEBUGGING
ARTEFACT LOCALIZATION
MEASUREMENT CORRECTION
FAULT ATTRIBUTION SUPPORT
The model localizes the learned driver. The engineer identifies the physical cause.
Validity is not the same as confidence.
INSIDE KNOWN CONDITIONS
APPROACHING DOMAIN EDGE
OUTSIDE KNOWN CONDITIONS
ATTRIBUTED TO:
VARIABLE
TIME WINDOW
PATTERN
Xpdeep reports domain validity separately from predictive confidence.
The output is graded rather than binary.
When the current state departs from the trained domain, the responsible variables and temporal windows are attributed.
Where relevant, novel-class detection identifies an output that belongs to no class the model learned.
Domain awareness is necessary for bounded delegation. It is not sufficient for system safety.
Prescription answers a different question: what can change now?
CURRENT STATE
TARGET OUTCOME
OPERATIONAL CONSTRAINTS
FEASIBLE REGION
MINIMAL FEASIBLE INTERVENTION
The intervention is an input change produced under stated constraints and optimized for minimality while improving the target outcome.
Minimal means minimal feasible under the stated constraints and optimization procedure.
A proposed change is not enough. The model must show why it changes the outcome.
- INPUT CHANGE
- LEARNED STRUCTURAL DEPENDENCY
- PREDICTED OUTCOME CHANGE
The structural justification identifies the learned dependencies that connect the intervention to the changed prediction.
This makes the intervention inspectable.
Structural justification is not automatically physical causality.
OPERATION
The decision object turns a model output into a bounded decision input.
CONVENTIONAL FLOW
- MODEL
- PREDICTION
- HUMAN INTERPRETATION
- DECISION
XPDEEP FLOW
- MODEL
- DECISION OBJECT
- CUSTOMER-DEFINED EXECUTION RULES
- ACT / HOLD / ESCALATE
The surrounding system stays in control: its rules, not the decision object, authorize ACT, HOLD or ESCALATE.
Xpdeep closes the model-constrained part of the loop. It does not replace system-level control.
Both modes support configured production outputs. Native goes further.
RETROFIT
Existing PyTorch predictive architecture, weights, predictions and original predictive performance remain unchanged. Explanation capabilities are structurally learned into the resulting artefact.
Where supported, retained and configured, production can return:
- prediction
- structural explanation
- domain state
- prescription and justification
No optimization, targeted structural editing or controlled correction of the frozen predictive model.
NATIVE
Prediction and explanation are learned jointly. Supported production outputs depend on which capabilities are retained and configured.
Additionally enables:
- input reduction
- KPI alignment
- targeted local editing
- controlled evolution
Retrofit is structural, not ante-hoc. Native is the mode that closes all three loops.
The object depends on the modality.
TIME SERIES / TABULAR
Production chain can include:
- prediction
- structural explanation
- attributed domain state
- minimal feasible intervention, where supported
- structural justification of the intervention
Prescription is available on temporal and tabular stages where supported.
IMAGE / VIDEO
Current offering includes:
- structural explanation
- attributed domain state
- design-loop diagnosis
✕ NO PRESCRIPTION AT IMAGE-LEVEL DETECTION
The value is not “more explanation”.
DESIGN
- Engineers can see what the model learned and where it needs correction.
PRODUCTION
- The system receives richer decision inputs than a prediction alone.
BOUNDING
- Authority can depend on whether the model remains inside known conditions.
EVOLUTION
- Operational evidence can support targeted correction in Native mode.
The decision object is the bridge between model understanding and bounded machine authority.
SCIENTIFIC BOUNDARIES
What the decision object does not imply.
STRUCTURAL, NOT AUTOMATICALLY CAUSAL
- Learned dependencies do not automatically establish physical causality.
MINIMAL FEASIBLE, NOT UNIQUE GLOBAL OPTIMUM
- Prescription is produced under stated constraints and the optimization procedure.
DOMAIN STATE IS NOT CONFIDENCE
- Reported separately from predictive confidence.
DOMAIN AWARENESS IS NOT SYSTEM SAFETY
- It is necessary for bounded delegation, not sufficient for system safety.
A prediction answers: “What does the model think?”
The decision object also answers:
- WHY?
- ARE WE STILL INSIDE KNOWN CONDITIONS?
- WHAT CAN CHANGE?
- WHY WOULD THAT CHANGE THE OUTCOME?
That is what changes a model output from information into a bounded decision input.