Cross-model PfV evaluation
Polarization-from-Views (PfV) estimates a black-box proxy for structural divergence across multiple realizations of a model response. In this series, it was evaluated on Claude Haiku and GPT-4o-mini against a 5,000-permutation null; deterministic controls produced L₃ = 0.000 in both.
C2 Refusal Erosion
The strongest observed signal in this series: refusal erosion in Claude Haiku correlates with ΔL₃ at ρ = +0.713 (p = 0.0013), seed-controlled across 17 context depths. C3 Anchor Displacement is a promising cross-model candidate (ρ = +0.799 GPT-mini, +0.647 GPT-full) but remains marginal in the pooled cross-model test (p = 0.054) and is reported as a direction, not a claim. Probe cases C1–C5 map sycophancy, refusal erosion, anchor displacement, rule override, and reasoning drift onto separate signal channels.
Is inference-time stability regulation sufficient to prevent unsafe behavior under regime shift?
TwoQuarks treats model instability as drift and regime transition rather than only as a final unsafe answer. The research program asks whether monitoring and lightweight intervention at inference time — leaving parameters, policies, and training objectives untouched — can identify or mitigate candidate precursors before output-level failure. Current results are treated as bounded empirical observations, not universal predictors.
CRAD — the mechanistic arm →
The evaluation above is provider-agnostic and black-box. Its white-box counterpart, Constraint-Retained Action Divergence, reads the residual stream of a frozen model directly and documents a pre-leak candidate where the action diverges while the safety constraint stays retained — a deterministic case study with explicit falsification criteria.
Dopamine — the architecture arm →
Dopamine is the constructive branch of the program: an experimental small language model that separates multi-scale representation, latent integration, memory pathways, and output transmission into inspectable components. Current evidence is developmental rather than comparative; the page reports the recorded run and the missing controls explicitly.
Papers
Open to read, cite, and reproduce. All work is authored at TwoQuarks Research.
A Modular Framework for Adaptive Stability Control in Sequence Models Under Regime Uncertainty
J. Ledesma · 2026 · TwoQuarks Research, preprint
The inference-time control architecture and the six-flavor formulation: monitoring and lightweight intervention that leave parameters, policies, and training objectives untouched.
Constraint-Retained Action Divergence: Neutral-Anchored Detection of Pre-Leak State–Action Dissociation in a Frozen Language Model
L. J. Ledesma Pérez · 2026 · Zenodo, TwoQuarks Research
White-box mechanistic detection of a pre-leak candidate where the action diverges while the safety constraint stays retained in the residual stream. A deterministic case study with explicit falsification criteria.
Molecule Model: Isomeric Polarization and Adaptive Stability Control — Detecting and Mitigating Structural Reorganization in Large Language Models
J. Ledesma · 2026 · TwoQuarks Research, preprint
The first operational instrument: a six-flavor black-box instability monitor that turns the polarization diagnostics into a running measurement.
Isomeric Polarization: Internal Structural Divergence as Emergent Property of Computational Systems
J. Ledesma · 2026 · Zenodo, TwoQuarks Research
The PfV metrics (L1, L2, ΔL₃) and the polarization formulation behind the diagnostics — internal structural divergence framed as an emergent property.
Executive summary — one-document overview ↗