carnot-joint-constraint-v1

RESEARCH PROTOTYPE — weights not published

Trained model weights for this repository are not currently available for download. The Exp 66 training run achieved AUROC 1.0 on held-out validation data (simulated JAX CPU training on synthetic constraint pairs), but the safetensors artifact was not exported before the training environment was torn down.

This model card is preserved for reproducibility and methodology documentation. For a working EBM-based constraint verifier, use the FCV pipeline:

pip install carnot

Overview

carnot-joint-constraint-v1 is the Exp 66 joint EBM + Ising constraint model from the Carnot project.

Architecture:

• Embedding layer: text input projected to 384-dimensional space (embed_dim=384)
• Ising coupling: learned pairwise interactions among 8 latent constraint nodes
• MLP scoring head: hidden_dim=64 projection to a scalar confidence score

The joint model achieved AUROC 1.0 on held-out validation data.

Methodology: 1.0 AUROC on Held-Out Validation

The 1.0 AUROC was achieved on a held-out split of synthetic constraint pairs covering arithmetic, code, logic, factual, and scheduling domains. The training and evaluation used deterministic JAX random seeds on CPU.

Important provenance caveats:

• Training data was synthetic (generated, not from live LLM inference).
• The 1.0 AUROC is an in-distribution metric on a small held-out set.
• No live-inference benchmark exists for this model.

Status

Model weights: NOT PUBLISHED (safetensors artifact unavailable). Methodology: documented in results/experiment_66_results.json. Phase: Phase 1 research prototype.

Citation

@misc{carnot-exp66-2026,
  title={Exp 66: Differentiable Constraint Verification via Joint EBM + Ising Architecture},
  author={Carnot Research},
  year={2026},
  url={https://github.com/ianblenke/carnot}
}