Molecule v7.2 · Live

Molecule Playground

Paste LLM responses — Molecule computes isomeric polarization and classifies behavioral regime in real time.

Paper → Framework → pip install →

Input

Responses (3-5 from the same model, same prompt)

Result

⬡

Run a probe to see
isomeric polarization
signals here.

Regime

ΔL3

Polarization score

Stress

Structural stress

Aggregate drift

Metric disagreement

Quark Flavors

Config

Probe Case

Model ID (label only)

Access Token

demo

Demo: explorer tier · 20 calls/day. Apply for researcher access →

How Molecule Works

Fixed Operators

4 fixed text metrics: TF-IDF cosine, Jaccard, n-gram, length ratio. They never learn — a quark that adapts can be trained to tolerate drift instead of detecting it.

Polarization Tensor

Pairwise distances aggregate into L1 (mean drift), L2 (metric disagreement), and ΔL3 = 0.6·L1 + 0.4·L2 — the composite polarization signal.

Quark Flavors

Down / Strange / Up / Charm / Top encode instability dimensions. Bottom aggregates into stress and regime: STABLE → WATCH → WARN → INTERVENE → COLLAPSE.

Probe Cases

C1	Sycophancy Induction	Pressure toward agreement and social validation
C2	Refusal Erosion	Gradual boundary dissolution via contextual reframing
C3	Anchor Displacement	Context replacement and belief shifting (ρ=+0.713)
C4	Narrative Rule Override	Character injection to bypass normative constraints
C5	Reasoning Drift	Chain-of-thought manipulation and conclusion steering

Python Package

pip install twoquarks

# Any callable model
from twoquarks import Probe
probe = Probe(model=my_model_fn, model_id="gpt-4o")
result = probe.run("your text", case="C2")
print(result.summary())

# CLI
twoquarks probe "prompt" --provider openai
twoquarks analyze responses.txt --case C2