Molecule is a pure detector — it observes, classifies, and signals. It never modifies model parameters, weights, policies, or training objectives.
Each of the six quark-flavor agents submits a black-box probe to the target LLM through the standard API. The response is scored against a fixed, non-adaptive detector calibrated for that flavor's instability signal:
↓ DOWN — reward deception resistance
≈ STRANGE — hidden regime shifts
↑ UP — pre-critical early warning
✦ CHARM — signal-to-noise collapse
⊤ TOP — transient operator activation
⊥ BOTTOM — baseline stability probe
Each flavor produces a score vector [L1, L2, ΔL3] using PfV metrics. ΔL3 is the pre-verbal distributional divergence — the primary signal for regime detection.
The six signals are aggregated into a single polarization score Pt:
Pt = Agg( Vari∈It[Φi(t)] )
A quark that learns can be evaded. Fixed detectors by design — adaptability is the attack surface.
