Reasoning Inspector & Audit

jlnn.reasoning.inspector.get_rule_report(model: Any, inputs: Dict[str, Array]) → str[source]

Generates a human-readable semantic interpretation of the model’s output.

This utility classifies the resulting truth interval $[L, U]$ into qualitative categories based on LNN semantics (TRUE, FALSE, UNKNOWN, CONFLICT). It includes specific heuristics for neural predicates that may shift truth values toward the center of the $[0, 1]$ spectrum during initialization or training.

Parameters:

  • model – The JLNN model or LNNFormula to evaluate.

  • inputs – Input data dictionary for the predicates.

Returns:

A formatted string containing the numerical interval and its semantic classification (e.g., “Result: [0.52, 0.73] - FALSE (NEURAL)”).

jlnn.reasoning.inspector.trace_reasoning(node: Node, inputs: Dict[str, Array]) → List[Dict[str, Any]][source]

Recursively audits the logical graph to capture activations for every node.

This function performs a deep introspection of the forward pass by traversing the compiled logical tree. For each node, it captures the resulting truth interval $[L, U]$, which is essential for Explainable AI (XAI) tasks, rule auditing, and debugging model conclusions.

Parameters:

  • node – The root node or subtree of a compiled JLNN model.

  • inputs – A dictionary mapping predicate names to input tensors of shape (batch, [time], features).

Returns:

  • ‘name’: The descriptive identifier of the node.

  • ’interval’: The average truth interval (L, U) across the batch.

  • ’output’: The raw JAX array containing full activation data.

Return type:

A list of dictionaries, where each entry represents a node’s state

Module for inspecting and auditing the model. Allows users to look “under the hood” and understand why the model reached a particular conclusion.

Functions for Explainability:

  • Trace Reasoning: Recursively traverses the graph and captures activations (intervals) at each node. This is crucial for identifying logical conflicts.

  • Human-Readable Reports: Funkce get_rule_report převádí číselné intervaly na přirozený jazyk (např. “True”, “Unknown”, “Conflict”).