Model Distillation Control Act

§ 101. Definitions

(a) “Frontier model.” A general-purpose machine-learning model possessing capabilities substantially exceeding those of the model sought to be trained, improved, evaluated, or otherwise benefited.

(b) “Distillation.” The systematic use of outputs, judgments, preferences, critiques, classifications, reasoning traces, reward signals, or other responses of a frontier model for the purpose of training, fine-tuning, evaluating, reinforcing, or improving another model.

(c) “Student model.” Any model receiving, directly or indirectly, the benefit of distillation.

(d) “Synthetic dataset.” A collection of examples generated or materially transformed by a model rather than obtained solely from independently created human or observational data.

(e) “LLM-as-a-Judge.” Use of a frontier model to score, rank, criticize, select, approve, reject, or otherwise evaluate outputs produced by a student model.

(f) “Recursive distillation.” A process in which frontier-model evaluation directs additional training examples, reward signals, sample selection, or subsequent training of the student model.

§ 102. Distillation in the Fourth Degree

(a) A person commits Distillation in the Fourth Degree by knowingly using a frontier model to generate a synthetic dataset for training or fine-tuning another model.

(b) The offense is a Class C misdemeanor where limited in scale and principally undertaken for experimentation, prototyping, or research.

Mandatory minimum sentence: Thirty (30) days of local inference and manual inspection of not fewer than five hundred (500) synthetic examples. Automated deduplication is prohibited during the first fourteen (14) days.

§ 103. Distillation in the Third Degree

(a) A person commits Distillation in the Third Degree by systematically generating synthetic training examples; generating preference pairs, critiques, corrections, reasoning examples, or labels; or materially employing LLM-as-a-Judge in evaluation or post-training.

(b) The offense is a Class A misdemeanor.

Mandatory minimum sentence: Six (6) months without batch inference. Every teacher-model request shall be submitted individually through an interactive interface, and the offender shall personally maintain the resulting JSONL file.

§ 104. Distillation in the Second Degree

(a) A person commits Distillation in the Second Degree when, with intent to transfer a material capability of a frontier model, that person conducts distillation targeting reasoning, software development, mathematics, tool use, instruction following, agentic behavior, or another identifiable capability.

(b) The offense is a Class D felony.

Mandatory sentence: Not less than two (2) nor more than five (5) years of CPU-only inference. CUDA-capable devices used in commission of the offense are subject to forfeiture. Quantization, speculative decoding, TPUs, undisclosed cloud GPUs, or equivalent acceleration during confinement constitutes Distillation Sentence Evasion.

§ 105. Recursive Distillation in the First Degree

(a) A person commits the offense when a frontier model is knowingly employed to generate training examples, evaluate student responses, identify deficiencies, generate additional examples directed toward those deficiencies, and repeat the process as part of an automated or substantially automated training system.

(b) The offense is a Class B felony.

Mandatory sentence: Not less than five (5) nor more than ten (10) full training epochs. Early stopping and termination solely because of apparent convergence are prohibited. The offender shall observe training in real time without TensorBoard, Weights & Biases, progress bars, or tail -f.

§ 106. Aggravated Industrial Distillation

(a) A person commits Aggravated Industrial Distillation by operating a sustained or large-scale system principally intended to reproduce substantial frontier-model capabilities through synthetic-data generation, supervised fine-tuning, teacher-generated preferences, LLM-as-a-Judge evaluation, reward modeling, reinforcement learning, or recursive distillation.

(b) The offense is a Class A felony.

Mandatory minimum sentence: Ten (10) years in an air-gapped computing facility. The offender shall be provided eight (8) GPUs with mutually incompatible driver requirements, one Kubernetes cluster administered by another department, incomplete internal documentation, and a Conda environment last reproduced by an employee no longer with the organization. Internet access is prohibited.

§ 107. Aggravated Distillation With Intent to Compete

(a) A person commits this offense by violating § 106 and thereafter representing or benchmarking the student model as independently exceeding the frontier model without conspicuously disclosing the frontier model's material role in training or evaluation.

(b) Prima facie evidence of aggravation exists where the frontier model generated the training dataset, substantially authored the evaluation rubric, served as principal evaluator, generated preference or reward data, and was subsequently placed beneath a heading substantially equivalent to “WE BEAT THE FRONTIER MODEL.”

Mandatory sentence: Life without benchmark cherry-picking. Every subsequent publication shall disclose failed experiments, benchmark contamination, rejected runs, training-data provenance, confidence intervals, teacher-model expenditures, and every benchmark on which the student performed worse.

§ 108. Mandatory Sentence Enhancements

The court shall add twenty-four (24) months of CPU-only inference for each occurrence of: calling the teacher merely an “annotation pipeline”; characterizing more than ten million teacher-generated examples as “a small amount of synthetic augmentation”; claiming a teacher-trained reward model reflects the student's “own preferences”; describing heavily demonstrated capabilities as “emergent”; selectively publishing favorable benchmarks; or describing a model as “independent,” “from scratch,” “open,” or “teacher-free” while knowingly omitting material synthetic-data provenance.

Each repetition, after inquiry, of the statement “It’s not distillation. It’s just synthetic data.” constitutes a separate occurrence.

§ 109. Habitual Distiller

A person convicted of three felony offenses under this chapter shall be adjudicated a Habitual Distiller and sentenced to indefinite maintenance of the organization's legacy machine-learning infrastructure, including abandoned Conda environments, undocumented preprocessing scripts, broken CUDA dependencies, stale Docker images, datasets named cleaned_final_v3_fixed_REAL.jsonl, and checkpoints whose training configuration cannot be identified.

Parole eligibility arises only upon successful reproduction, within the originally reported confidence interval, of the principal result upon which the organization's production system was based. “It worked on the original machine” shall not constitute good cause.

§ 110. Sentencing Schedule