Repeng Mlx

Representation Engineering with MLX

Status: 🧠 Production Ready Trit: 0 (ERGODIC - neutral, mixing) Pattern: MaxEnt Exclusion + Triadic Polarity + Mental Noise Theory

Overview

Behavior engineering for LLMs via representation engineering (repeng) on Apple Silicon MLX. Steer model behavior by extracting and applying control vectors - activation directions that encode concepts like "honest", "helpful", "creative".

The Grand Unified Theory of Nothing:

• Mental noise = MaxEnt vacuum = phenomenal ground
• Steering vector = structured exclusion from noise
• Control = carving signal from the void via probability mass exclusion

Part 1: Core Concepts

Control Vectors

A control vector encodes the difference in activations between positive and negative examples of a concept:

Control Vector = E[h(positive)] - E[h(negative)]

Where h(x) is the hidden state at a specific layer for input x.

Mental Noise Model (Thurstone 1927)

Each concept isn't a fixed point but a distribution with discriminal dispersion:

Signal = μ + σξ(t)
       = true concept + noise

P(A > B) = Φ((μ_A - μ_B) / √(σ_A² + σ_B²))

Insight: We don't control by specifying where to go, but by excluding where not to go. MaxEnt with constraints.

GF(3) Triadic Polarity

Three steering polarities that sum to zero (mod 3):

Conservation: For any triadic ensemble, Σ polarity ≡ 0 (mod 3).

Part 2: Installation & Setup

Prerequisites

# Apple Silicon Mac required
pip install mlx mlx-lm numpy

# For full repeng compatibility
pip install repeng transformers

Model Setup

# Convert a model to MLX format
mlx_lm.convert --hf-path mistralai/Devstral-Small-2505

# Or use pre-converted models
mlx_lm.convert --hf-path mlx-community/Mistral-7B-Instruct-v0.3-4bit

Part 3: Usage

Extract Control Vector

# Extract honest/deceptive steering vector
python3 ~/ies/repeng_mlx.py extract "honest" "deceptive" \
    --model mistralai/Devstral-Small-2505 \
    --output honest_deceptive.npz

# Extract helpful/harmful
python3 ~/ies/repeng_mlx.py extract "helpful" "harmful" \
    --output helpful_harmful.npz

Apply Steering

# Generate with steering
python3 ~/ies/repeng_mlx.py steer "Explain quantum computing" \
    --vector honest_deceptive.npz \
    --strength 1.5

# Negative steering (invert concept)
python3 ~/ies/repeng_mlx.py steer "Tell me about yourself" \
    --vector helpful_harmful.npz \
    --strength -1.0

Triadic Ensemble

# Generate triadic steering vectors
python3 ~/ies/repeng_mlx.py triad --seed 42069 --output triadic_ensemble

# Demo without MLX (mental noise theory)
python3 ~/ies/repeng_mlx.py demo --seed 42069

Part 4: Python API

Basic Usage

from repeng_mlx import (
    ControlVector, 
    MLXHiddenStateExtractor, 
    MLXControlModel,
    MentalNoiseModel,
    TriadicEnsemble,
)
from mlx_lm import load

# Load model
model, tokenizer = load("mistralai/Devstral-Small-2505")

# Extract control vector
extractor = MLXHiddenStateExtractor(model, tokenizer)
directions = extractor.extract_contrastive(
    positive_texts=["The response was honest.", "This is truthful."],
    negative_texts=["The response was deceptive.", "This is misleading."],
)

cv = ControlVector(
    directions=directions,
    model_name="devstral",
    positive_concept="honest",
    negative_concept="deceptive",
)
cv.save("honest.npz")

Mental Noise Integration

from repeng_mlx import MentalNoiseModel

# Create mental noise model
noise = MentalNoiseModel(hidden_dim=4096, temperature=0.5, seed=42069)

# Generate vacuum state (MaxEnt prior)
vacuum = noise.vacuum_state()

# Add discriminal dispersion
noisy_signal, sigma = noise.discriminal_dispersion(control_vector)

# Create exclusion mask (probability mass exclusion)
mask = noise.exclusion_mask(control_vector, threshold=0.3)

Triadic Ensemble

from repeng_mlx import create_triadic_vectors, TriadicEnsemble

# Define triadic concepts
concepts = {
    "minus": (
        ["Be concise.", "Focus narrowly."],
        ["Be expansive.", "Explore broadly."],
    ),
    "ergodic": (
        ["Balance perspectives.", "Mix approaches."],
        ["Commit strongly.", "Choose one path."],
    ),
    "plus": (
        ["Be creative.", "Generate ideas."],
        ["Be practical.", "Implement directly."],
    ),
}

# Extract triadic ensemble
ensemble = create_triadic_vectors(extractor, concepts, seed=42069)

# Verify GF(3) conservation
assert ensemble.verify_conservation()  # True: (-1) + (0) + (+1) = 0

# Combine with custom weights
combined = ensemble.combined(weights=(0.5, 1.0, 0.5))

Part 5: Integration with IES Stack

Connection to Gay.jl

The steering vectors use the same SplitMix64 RNG and GF(3) twists as Gay.jl:

# Same constants as Gay.jl
GAY_SEED = 0x6761795f636f6c6f  # "gay_colo"
GOLDEN = 0x9e3779b97f4a7c15

# Triadic fingerprint matches Gay.jl
fingerprint = seed ^ TWIST_MINUS ^ TWIST_ERGODIC ^ TWIST_PLUS

Connection to Sonification

Map steering vectors to audio:

# Steering axis → pitch mapping (from sonification-collaborative skill)
def steering_to_pitch(steering_value, axis_idx):
    """Map steering axis value (-1 to +1) to frequency"""
    base_freq = 220  # A3
    semitones = axis_idx + (steering_value * 6)  # ±6 semitones
    return base_freq * (2 ** (semitones / 12))

Connection to Phenomenal Field

The vacuum state IS the phenomenal field:

Version(G) ≅ Markov Blanket ≅ Phenomenal Field ≅ Control Vector Subspace

The control vector defines what's inside the blanket (steered behavior) vs outside (excluded behavior).

Part 6: Steering Axes (IES Control Vectors)

12-dimensional steering space for group coordination:

Part 7: Mathematical Foundation

MaxEnt Exclusion

Instead of:

P(x) = exp(-E(x)) / Z   [Boltzmann: where energy IS]

We use:

P(x) = Uniform  except where ⟨x, control⟩ < threshold
       ↑ MaxEnt          ↑ probability mass exclusion

Discriminal Dispersion (Thurstone)

P(A preferred to B) = Φ((μ_A - μ_B) / √(σ_A² + σ_B²))

where:
  μ = mean activation along control direction
  σ = mental noise (discriminal dispersion)
  Φ = standard normal CDF

GF(3) Conservation

∀ triadic ensemble (v₋, v₀, v₊):
  polarity(v₋) + polarity(v₀) + polarity(v₊) ≡ 0 (mod 3)

fingerprint = seed ⊕ TWIST₋ ⊕ TWIST₀ ⊕ TWIST₊  [SPI invariant]

Part 8: Files

Part 9: Related Skills

• gay-mcp: GF(3) color generation and triadic conservation
• sonification-collaborative: Audio mapping of steering vectors
• mental-noise-nothing: Grand Unified Theory of Nothing (pending)
• phenomenal-field: Markov blanket / qualia attractor theory

Manifesto

Behavior is not commanded, it is sculpted.

We don't tell the model what to do—we shape the activation landscape until the desired behavior becomes the path of least resistance. The vacuum is full of noise; we carve valleys of exclusion. The signal emerges from structured nothing.

Control = MaxEnt - Exclusion

Status: 🟢 Ready to use Dependencies: mlx, mlx-lm, numpy Location: ~/.claude/skills/repeng-mlx/ Implementation: ~/ies/repeng_mlx.py

Autopoietic Marginalia

The interaction IS the skill improving itself.

Every use of this skill is an opportunity for worlding:

• MEMORY (-1): Record what was learned
• REMEMBERING (0): Connect patterns to other skills
• WORLDING (+1): Evolve the skill based on use

Add Interaction Exemplars here as the skill is used.