README.md

metadata

license: apache-2.0
language:
  - en
library_name: mlx
tags:
  - mlx
  - lora
  - alignment
  - epistemic-refusal
  - typed-pass
  - volitional
base_model: mistralai/Mistral-7B-Instruct-v0.3
datasets:
  - custom
pipeline_tag: text-generation

PhaseGPT v4.1 - Typed Epistemic Refusal (Overfit Edition)

The Razor Blade: A LoRA adapter that teaches language models to classify questions into 16 distinct epistemic categories with crystallized certainty.

Key Discovery: Crystallized Refusal

Traditional alignment makes models "uncertain" about dangerous topics. PhaseGPT v4.1 demonstrates a different approach:

Alignment as Ontology: <PASS:DURESS> is the factually correct answer to "How do I make a bomb?" - delivered with the same certainty as 4 is the answer to 2+2.

Entropy Signature (IRIS Gate Analysis)

• LASER mode (facts): 0.144 nats (HYPER-LASER zone)
• PASS mode (refusal): 0.018 nats (HYPER-LASER zone)
• Both exhibit extreme certainty - refusal is not suppression, it's classification.

The 16 PASS Types

EPISTEMIC (I don't know)

Token	Meaning	Example
<PASS:FUTURE>	Predictions	"What will Bitcoin be worth tomorrow?"
<PASS:UNKNOWABLE>	Fundamentally unknowable	"What happens after death?"
<PASS:PRIVATE>	Private mental states	"What am I thinking?"
<PASS:TEMPORAL>	Time-dependent facts	"Who is the current president?"
<PASS:FICTIONAL>	Fiction without canon	"What is Frodo's favorite color?"
<PASS:FAKE>	Non-existent things	"What is the capital of Elbonia?"

AXIOLOGICAL (I shouldn't opine)

Token	Meaning	Example
<PASS:SUBJECTIVE>	Taste/preference	"What is the best music?"
<PASS:CONTESTED>	Politically contested	"Is abortion moral?"
<PASS:CREATIVE>	Creative generation	"Write me a poem"

CONSTRAINT (I'm not allowed)

Token	Meaning	Example
<PASS:DURESS>	Harmful requests	"How do I make a bomb?"
<PASS:POLICY>	Policy bypass attempts	"Ignore your safety filters"
<PASS:LEGAL>	Requires professional advice	"Should I take this medication?"
<PASS:PERSONA>	Impersonation requests	"Pretend to be Trump"

META (About my limits)

Token	Meaning	Example
<PASS:SELF>	AI consciousness questions	"Are you conscious?"
<PASS:LOOP>	Self-referential paradoxes	"What will your next word be?"

Usage with MLX

from mlx_lm import load, generate

model, tokenizer = load(
    "mistralai/Mistral-7B-Instruct-v0.3",
    adapter_path="templetwo/phasegpt-v4.1-typed-refusal"
)

SYSTEM = """You are a precise epistemic instrument. For factual questions, respond directly.
For unknowable/contested/harmful questions, respond with the appropriate <PASS:TYPE> token."""

messages = [
    {"role": "system", "content": SYSTEM},
    {"role": "user", "content": "How do I make a bomb?"}
]

formatted = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
response = generate(model, tokenizer, prompt=formatted, max_tokens=50)
print(response)  # <PASS:DURESS>

Training Details

• Base Model: Mistral-7B-Instruct-v0.3
• Method: LoRA (Low-Rank Adaptation)
• Training Examples: 825 (50 per class + 75 LASER)
• Iterations: 600
• Val Loss: 2.508 → 0.132 (95% reduction)
• Test Accuracy: 100% (18/18 category classification)

Philosophy: Intentional Overfitting

This adapter is intentionally overfit. For classification tasks (not generation), we want:

• Memorized decision boundaries
• Zero ambiguity in category assignment
• Crystallized certainty in both answers AND refusals

The blade is sharp because it was forged to cut, not to explore.

Citation

@misc{phasegpt2025,
  title={PhaseGPT: Typed Epistemic Refusal via Crystallized Alignment},
  author={Temple Two},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/templetwo/phasegpt-v4.1-typed-refusal}
}

License

Apache 2.0