A working model of collapse as inference-time instability — not only capacity.
TwoQuarks models collapse as a potential failure of inference-time stability rather than capacity alone, and explores a modular control layer. Six modular flavors monitor candidate pre-instability signals and apply targeted interventions during execution — operating entirely at inference time — while 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 a black-box proxy for structural divergence from API outputs alone. The composite signal ΔL₃ summarizes divergence across realization sets and is used to test whether such changes emerge before output-level failure.
Graph Momentum
A graph-temporal extension for measuring oscillation, resistance, and scalar/graph dissociation — dynamics the scalar ΔL₃ trajectory alone may not capture.
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.