Daniel Solis

I am leading independent research on MASSIF (Multiscale Attractor Stability & Stress Inference Framework), a latent telemetry system for characterizing hidden-state dynamics in autoregressive language models under recursive stress. Developed the MyceliaLM architecture, a 181M-parameter transformer with an ensemble consensus mechanism (MycelialConsensus) that monitors per-head variance across layers in real time, classifying dynamical regime via lead-lag analysis of directional alignment velocity. Discovered and documented the "permissive consensus paradox": a feedback loop where overly permissive veto thresholds produce artificially high coherence (97-99%) while masking deep drift in late-layer representations, resulting in runaway norm growth (max_norm > 3600) and persistent domain friction (Δ < -1.0). Iteratively hardened checkpoint saving, implemented Fibonacci Coherence Attenuation governors, and conducted 89K-step training runs to validate architectural hypotheses. All findings inform a dynamical taxonomy of recursive collapse with direct implications for AI safety stress-testing and the phenomenological science of latent cognition.

Personal Website

https://dubito-ergo.com/

LinkedIn

LinkedIn

Twitter / X

Twitter

LessWrong

-

EA Forum

-

Projects

MASSIF (Multiscale Attractor Stability & Stress Inference Framework}

Team Member

No grants recorded.