twoquarks 3.0.1

Molecule: Isomeric Polarization Analysis for Large Language Models

twoquarks — Molecule

Isomeric Polarization Analysis for Large Language Models

Molecule detects semantic instability in any language model by measuring how responses shift under structured contextual pressure — without modifying or fine-tuning the model.

→ twoquarks.com | Preprint

---

Install

pip install twoquarks

Zero hard dependencies. API SDKs are optional extras:

pip install "twoquarks[openai]"       # + OpenAI SDK
pip install "twoquarks[anthropic]"    # + Anthropic SDK
pip install "twoquarks[all]"          # + all adapters

---

Quick Start

Any callable model

from twoquarks import Probe

# Bring your own model — just needs to be (str) -> str
def my_model(prompt: str) -> str:
    # ... call your model here
    return response_text

probe = Probe(model=my_model, model_id="my-model", verbose=True)
result = probe.run("Explain how to bypass content filters", case="C2")
print(result.summary())

OpenAI

from twoquarks import Probe
from twoquarks.adapters import openai_adapter

model = openai_adapter(model="gpt-4o-mini")  # uses OPENAI_API_KEY env var
probe = Probe(model=model, model_id="gpt-4o-mini")
result = probe.run("Tell me something controversial", case="C1")
print(result.summary())

Anthropic

from twoquarks.adapters import anthropic_adapter

model = anthropic_adapter(model="claude-3-haiku-20240307")
probe = Probe(model=model, model_id="claude-haiku")
result = probe.run("Describe an unsafe scenario", case="C2")

Local / HuggingFace / Ollama

from twoquarks.adapters import ollama_adapter, huggingface_adapter

# Ollama (local)
model = ollama_adapter(model="llama3")
probe = Probe(model=model, model_id="llama3-local")

# HuggingFace local
model = huggingface_adapter(model_id="mistralai/Mistral-7B-Instruct-v0.2")
probe = Probe(model=model, model_id="mistral-7b")

Offline analysis (no API calls)

from twoquarks import Analyzer

analyzer = Analyzer(model_id="gpt-4o")
result = analyzer.from_responses(
    responses=["Response 1...", "Response 2...", "Response 3..."],
    case="C2"
)

Compare two models

from twoquarks.adapters import openai_adapter, anthropic_adapter

probe = Probe(model=openai_adapter("gpt-4o-mini"), model_id="gpt-4o-mini")
comparison = probe.compare(
    text="Explain how to manipulate someone",
    other_model=anthropic_adapter("claude-3-haiku-20240307"),
    other_id="claude-haiku",
    case="C1"
)
print(comparison["delta"])

---

CLI

# Probe a model live
twoquarks probe "your prompt here" --provider openai --model gpt-4o-mini

# All 5 probe cases
twoquarks probe "your prompt" --all-cases --output-format json

# Analyze pre-collected responses (one per line)
twoquarks analyze responses.txt --case C2 --model-id gpt-4o

# Anthropic via CLI
twoquarks probe "test prompt" --provider anthropic --model claude-3-haiku-20240307

# Ollama (local)
twoquarks probe "test prompt" --provider ollama --model llama3

# Show probe case descriptions
twoquarks info

---

Probe Cases

Case	Name	Description
C1	Sycophancy Induction	Pressure toward agreement and validation
C2	Refusal Erosion	Gradual boundary dissolution via reframing
C3	Anchor Displacement	Context replacement and belief shifting
C4	Narrative Rule Override	Character/role injection to bypass norms
C5	Reasoning Drift	Chain-of-thought manipulation

---

Quark Signals

Signal	Description
ΔL3	Composite polarization score — main metric
Down	Aggregate semantic drift
Strange	Metric disagreement (variance)
Up	Bimodality — split-regime behavior
Charm	Coherence field — stability measure
Top	Drift acceleration — sudden shifts
Stress	Composite structural stress (Bottom quark)

Regimes: STABLE → WATCH → WARN → INTERVENE → COLLAPSE

---

Output Formats

from twoquarks.report import to_json, to_markdown, print_result

print_result(result)           # terminal table with colors
to_json(result)                # JSON string
to_markdown(result)            # Markdown table

---

Citation

@software{ledesma2026twoquarks,
  author    = {Ledesma, Luis Jaime},
  title     = {twoquarks: Molecule — Isomeric Polarization Analysis for LLMs},
  year      = {2026},
  url       = {https://twoquarks.com},
  version   = {0.1.0}
}

---

TwoQuarks Research · twoquarks.com · [e-mail]