Retraining vs Reinduction

The practical distinction

Retraining asks: "Can the existing model structure learn the updated behaviour?" Reinduction asks: "Does the problem now require a different structure?"

The distinction matters because an XNN contains more than weights. Its modules identify important features and interactions; its partitions define local behaviour; and its rules create the explanation scaffold. Updating parameters inside an unsuitable scaffold can preserve apparent continuity while failing to represent the new problem.

Retraining preserves structure

Retraining is appropriate when:

• the input and target schema remain compatible;
• the same modules and interactions remain meaningful;
• partition boundaries still provide adequate coverage;
• change is gradual or calibrational; and
• monitoring does not reveal persistent new activation patterns that the structure cannot represent.

The coefficients in the local expressions may change, and resulting attributions may shift. The important point is that the existing structure remains a valid basis for learning and explanation.

Reinduction creates a new structure

Reinduction should be considered when:

• input features are added, removed or redefined;
• the predictive target changes;
• new categories introduce materially different behaviour;
• a business model or operating process changes substantially;
• new data collection technology changes the meaning of inputs;
• a new causal or policy structure must be represented; or
• existing modules and partitions no longer cover the observed relationships.

Reinduction repeats dataset analysis, interaction discovery, partition formation, rule generation and neuro-symbolic packaging.

Market disruption is not an automatic answer

A major event may require retraining, reinduction or a temporary operational control. The choice depends on evidence.

If the disruption changes coefficients within known relationships, retraining may be sufficient. If it creates new interactions or invalidates assumptions, reinduction may be necessary. If data is too limited to establish either, the organisation may use a controlled fallback or review path while evidence accumulates.

A decision framework

Evaluate five areas:

1. Schema: are the fields, categories, units and targets unchanged?
2. Coverage: are incoming queries using established partitions, or are unfamiliar paths growing?
3. Behaviour: have attribution shapes and rule effects shifted within the existing structure?
4. Knowledge: have policies, causal assumptions or expert rules changed?
5. Objective: is the model still solving the same operational problem?

Minor changes in one area may support retraining. Material changes across several areas point towards reinduction.

Comparing the options

Retraining can preserve stable identifiers and make parameter-level comparison direct. It may be faster and less disruptive where the structure remains valid.

Reinduction allows the system to discover new modules, partitions and interactions. It also creates a larger assurance task because structural explanations can change. A new module may become dominant, a partition boundary may move or a human-injected rule may need to be remapped.

The assurance plan should reflect this difference. Retraining review can focus on parameter, attribution and outcome changes within a known structure. Reinduction review must additionally justify the new structure and confirm how earlier controls, tests and symbolic references were transferred or replaced.

Evidence before deployment

Both paths require evaluation. Compare predictive metrics, calibration, subgroup performance, attribution patterns, activation coverage and explanation stability. Confirm that human knowledge and operational policy are represented as intended.

For reinduction, structural comparison should identify added, removed and changed modules and rules. For retraining, parameter and attribution comparison should verify that preserved identifiers still refer to equivalent components.

Version lineage

Treat retraining and reinduction as different relationships in the model registry. A retrained model is a new version of an existing structure. A reinduced model is a new structure with lineage to the prior system and problem definition.

This distinction helps reviewers understand whether behavioural continuity is expected and which comparisons are meaningful.

Why it matters

Choosing the wrong path can either create unnecessary model churn or preserve an obsolete explanation structure. A transparent decision framework connects lifecycle action to observed change.