Add LNNClassifier for IBM LNN integration

[ColtonPayne]

Summary

Introduces LNNClassifier, a bidirectional bridge between IBM Logical Neural Networks and PyReason. This enables a feedback loop where:

• LNN reasons locally about individual entities using learned logical rules over bounded truth values [lower, upper]
• PyReason propagates globally across graph structure using its rule engine
• Feedback tightens bounds — PyReason's graph-level results feed back into the LNN as new evidence, producing results neither system achieves alone

Why this matters

LNN and PyReason both operate on interval-valued truth bounds [L, U], making them natural integration partners. But they have complementary blind spots:

Capability	LNN	PyReason
Local logical composition (e.g. expert ∧ endorsed → reliable)	Yes	No
Graph-structure propagation (e.g. traverse trust edges)	No	Yes

The LNNClassifier bridges this gap, enabling multi-hop reasoning that combines local logical inference with global graph propagation.

Verified with POC

Check it out here

Changes

New files

• pyreason/scripts/learning/classification/lnn_classifier.py
  • LNNClassifier — wraps an IBM LNN Model; forward() runs inference and emits PyReason Fact objects; receive_feedback() injects PyReason bounds back as LNN evidence
  • LNNInterfaceOptions — configures convergence_threshold, max_feedback_rounds, bound_tightening_only

Modified files

• pyreason/pyreason.py — conditional import of LNNClassifier / LNNInterfaceOptions when both torch and lnn packages are installed (graceful fallback to None otherwise)

Design decisions

• Does NOT extend LogicIntegrationBase — LNN Model is not a torch.nn.Module, but LNNClassifier follows the same forward(data, t1, t2) → (raw, postproc, List[Fact]) interface contract
• Bound tightening only (default) — feedback from PyReason only tightens LNN bounds, never widens them, preserving logical consistency
• predicate_map bridges LNN's stateful Predicate objects to string names; node_groundings maps LNN grounding names to PyReason node names
• target_predicates controls which LNN predicates export as PyReason Facts (not all need exporting)

Dependencies

• Requires lnn package: pip install git+https://github.com/IBM/LNN
• Optional — falls back gracefully when not installed

Test plan

• Pre-commit hooks pass (ruff, unit tests with JIT enabled/disabled)
• Trust network POC verified end-to-end (3 scenarios: LNN-only, PyReason-only, combined)
• Graceful fallback when lnn not installed (LNNClassifier = None)

🤖 Generated with Claude Code