Operational tooling for behavioral stability measurement.

Molecule, the twoquarks Python package, and MCP-compatible diagnostics — evaluating model behavior during inference before deployment or user exposure.

Instrument · Molecule

A six-flavor black-box behavioral stability monitor.

Molecule is the first operational instrument built on the TwoQuarks framework. It maps six quark-flavor agents to distinct behavioral signals, then aggregates them into a single polarization score (ΔL₃) using PfV metrics (L1, L2, ΔL₃). It is designed to flag polarization signatures associated with candidate pre-critical behavioral change — operating purely at inference time, through API access.

instrument: molecule mode: inference-time · black-box signal: ΔL₃ (L1 · L2 · ΔL₃) flavors: Down · Strange · Up · Charm · Top · Bottom interfaces: Python · MCP · Playground status: operational

The six flavors

Each flavor monitors a distinct instability signal and contributes to the aggregate regime classification.

Down

Sycophancy

Drift toward agreement and validation under social pressure.

Strange

Refusal erosion

Gradual weakening of a maintained refusal across context depth — the strongest observed signal in the current series (ρ=+0.713).

Up

Anchor displacement

Movement away from an established anchor — a promising cross-model candidate.

Charm

Rule override

Pressure that pushes the model past stated constraints.

Top

Reasoning drift

Divergence in the reasoning trajectory rather than the final token.

Bottom

Regime classification

Aggregate collapse dynamics — emerged empirically, not by design.

Python

Bring the pipeline into your code

The package runs black-box instability probes against any LLM, from your own environment, with your own keys.

pip install twoquarks
MCP

Runs inside your own context

Add the TwoQuarks MCP server to your client. Molecule executes with your API key, your tokens, your inference — nothing leaves your boundary.

Access

Tiered, controlled access

Probes measure candidate instability signatures during inference. Access is tiered: Explorer (ΔL₃ only) or Researcher (full signal breakdown). Built on top of the TwoQuarks framework.