Carnot-EBM/carnot-constraint-propagation-logic

Research Artifact — Not Production-Ready This model verifies logical syllogism responses using structural binary features. It achieves AUROC 1.0 on the held-out test set (matching Exp 89 reference). It handles modus ponens, modus tollens, disjunctive syllogism, and affirming the consequent — not arbitrary logic or modal reasoning.

constraint-propagation-logic

Learned Ising constraint model for logical syllogism verification.

Trained via discriminative Contrastive Divergence (Exp 62/89) on 400 verified (question, correct, wrong) triples for four syllogism types. Assigns a scalar energy to binary-encoded responses — lower energy means the response correctly identifies whether a syllogism is valid or invalid.

What It Is

An Ising Energy-Based Model (EBM):

E(x) = -(b^T s + s^T J s)    s = 2x - 1 ∈ {-1, +1}^200

The coupling matrix J encodes feature pair co-occurrences that distinguish valid from invalid logic responses (e.g., "follows by modus ponens" + "this follows" co-occurring is a strong correct-response signal).

How It Was Trained

Algorithm: Discriminative Contrastive Divergence (full-batch)

Training data: Programmatically generated syllogism triples — four types:

1. 1.Modus ponens (valid): "If all X are P, and all Y are X, what follows?" → "All Y are P. This follows by modus ponens."
2. 2.Modus tollens (valid): "If all X are P, and Y are not P, what follows?" → "Y are not X. This follows by modus tollens."
3. 3.Disjunctive syllogism (valid): "Either X are P or X are Q. X are not P. What follows?" → "X are Q. This follows by disjunctive syllogism."
4. 4.Affirming the consequent (invalid): "If all X are P, and Y are P, are Y necessarily X?" → "No. This is affirming the consequent."

Wrong answers reverse the conclusion or deny the correct inference pattern.

Benchmark Results

Perfect separation: the structural features of valid-logic responses (keywords "follows by", "modus ponens", "modus tollens", "disjunctive syllogism", "affirming the consequent") are highly discriminative.

Usage

import numpy as np
from carnot.inference.constraint_models import ConstraintPropagationModel

# Load model
model = ConstraintPropagationModel.from_pretrained(
    "exports/constraint-propagation-models/logic"
)

# Encode a response
from scripts.export_constraint_models import encode_answer
question = "If all cats are mortal, and all robots are cats, what follows?"
answer = "All robots are mortal. This follows by modus ponens."
x = encode_answer(question, answer)

energy = model.energy(x)  # low = valid syllogism
score = model.score(x)    # high = more likely a correct logic response
print(f"Energy: {energy:.2f}, Score: {score:.3f}")

Limitations

1. 1.Keyword-dependent: Perfect AUROC is achieved because valid syllogism responses contain distinctive keywords ("modus ponens", "follows by"). A correct response without these keywords may score poorly.
2. 2.Four syllogism types only: Trained on modus ponens, modus tollens, disjunctive syllogism, and affirming the consequent. Other logic forms (hypothetical syllogism, constructive dilemma) are out of scope.
3. 3.Template-generated training data: Real LLM responses vary in phrasing and may not contain the expected keyword cues.
4. 4.Feature encoder required: Must use the same 200-dim structural encoder.

Files

Citation

@misc{carnot2026constraint_logic,
  title  = {Carnot Constraint Propagation Model: Logic},
  author = {Carnot-EBM},
  year   = {2026},
  url    = {https://github.com/ianblenke/carnot}
}

Spec

• —REQ-VERIFY-002, REQ-VERIFY-003, FR-11

Production Use

For production LLM output verification, install: pip install carnot

This model is a Phase 1 research artifact (activation-based confidence detection). These 16 per-token activation EBMs detect confidence, not correctness. For the full verify-repair pipeline, see: https://github.com/Carnot-EBM/carnot-ebm