jepa-predictor-v2

JEPA Violation Predictor v2 — a multi-domain neural verifier trained to predict whether a language model answer violates domain constraints (arithmetic, code, logic).

This is a research artifact from the Carnot EBM framework. It is a small MLP (input_dim=200, hidden_dim=128, output_dim=1) trained with class-weighted binary cross-entropy and early stopping on a balanced multi-domain dataset.

Performance (Exp 155)

Domain	AUROC (v2)	vs v1
Arithmetic	0.721	+0.018
Code	0.776	+0.071
Logic	0.479	−0.056
Macro average	0.659	+0.011

Training details: 963 samples, best epoch 19 / 100, early stopping on validation macro AUROC (patience=15), class-weighted BCE loss (pos_weight clipped [0.1, 10]).

Architecture

JEPAViolationPredictor(
  input_dim   = 200    # 200-dim binary feature vector (same encoder as Ising models)
  hidden_dim  = 128
  output_dim  = 1      # P(violation | x)
)

The predictor is trained on structured (question, answer, violated) triples. Features are the same 200-dim binary structural encoding used by the Ising constraint models (Carnot-EBM/constraint-propagation-*). This makes the two model families interoperable.

Training Data (Exp 154 → Exp 155)

• Arithmetic: 800 pairs reused from Exp 143 (carry-chain arithmetic templates)
• Code: 200 synthetic pairs (Python type, return, initialisation constraints)
• Logic: 200 synthetic pairs (implication, exclusion, disjunction, negation)
• Total: 1,200 pairs; 963 train / 237 validation (stratified by domain × violated)

Usage

from safetensors.numpy import load_file

weights = load_file("jepa_predictor_v2.safetensors")
# Keys: "layer1.weight", "layer1.bias", "layer2.weight", "layer2.bias"

# Or load via the carnot package:
from carnot.inference.jepa_predictor import JEPAViolationPredictor
predictor = JEPAViolationPredictor.load("jepa_predictor_v2.safetensors")

# Score a 200-dim binary feature vector
import numpy as np
x = np.zeros(200, dtype=np.float32)
x[:20] = 1.0   # example features
prob_violation = predictor.predict(x)
print(f"P(violation) = {prob_violation:.3f}")

Limitations

1. Logic AUROC low (0.479): byte-level structural features do not capture logical implication structure well. Semantic embeddings are needed.
2. Code domain fast-path issues (Exp 156): at threshold ≥ 0.3, all code questions are fast-pathed (200/200), causing degradation. Use threshold ≥ 0.8 for code.
3. Target not met: <2% degradation target at t=0.5 not achieved. Treat as directional baseline, not production-ready.
4. Template training data: Not validated on real LLM outputs.

Note: Production Installation

Note: This is a Phase 1 research artifact. For production use of the full Carnot EBM framework (constraint verification, guided decoding, energy-based repair), see:

pip install carnot

Source and documentation: https://github.com/ianblenke/carnot

Spec

• REQ-JEPA-001: Violation predictor trained on multi-domain data.
• REQ-JEPA-002: Fast-path routing using predictor confidence.
• SCENARIO-JEPA-003: Cross-domain AUROC improvement over v1.

Citation

@misc{carnot2026jepa,
  title  = {Carnot JEPA Violation Predictor v2},
  author = {Carnot-EBM},
  year   = {2026},
  url    = {https://github.com/ianblenke/carnot}
}