Model collapse as a failure of inference-time stability — not capacity.
TwoQuarks defines collapse as a failure of inference-time stability rather than capacity, and introduces a modular control layer. Six modular flavors monitor pre-instability signals and apply targeted interventions during execution — operating entirely at inference time — model parameters, policies, and training objectives stay untouched. Six flavors, one coherent system.
PfV & ΔL₃
PfV uses multiple realizations of a model response to estimate structural divergence from API outputs alone. The composite signal ΔL₃ measures divergence across realization sets and detects instability before the final output fails.
Graph Momentum
A graph-temporal extension for detecting oscillation, resistance, and scalar/graph dissociation — instability the scalar ΔL₃ trajectory alone can hide.
C1–C5: behavioral failure modes
The framework maps distinct failure modes onto specialized signal channels: sycophancy, refusal erosion, anchor displacement, rule override, and reasoning drift — with Bottom providing aggregate regime classification. The Bottom component emerged empirically from observed collapse, not from theoretical design.