Publication Program

Below a family-dependent critical scale N_c, the coupling between reasoning (HellaSwag) and truthfulness (TruthfulQA) is negative — scaling reasoning hurts truthfulness. Above N_c, they cooperate. N_c varies 60× across families (0.12B–7B) and is a design parameter, not a physical constant: width, data curation, and architecture each shift it independently. A coupled ODE cross-predicts held-out Llama-2 at 5.6% MAE. The isocline classifier separates standard-trained from curated families. Curated models (Phi, Qwen3) bypass the tax entirely.

At frontier scale (SWE-bench vs GPQA Diamond, 34+5 models, 10 labs), capabilities remain cooperative (r = +0.72, slope 0.513). The h-field diagnostic — deviation from the cooperation trend — reveals each lab’s training philosophy: Google is reasoning-specialist (h̄ = +5.5), Anthropic is coding-rich (h̄ = −6.9). Per-lab coupling slopes span 5× (Google 1.15 vs DeepSeek 0.23). Tax excursions are temporary — Sonnet 4.6 (h = −13.1) recovers at Opus 4.6 (h = +3.5). The h-field is descriptive, not causal. Seven falsifiable predictions with timestamped deadlines.

Agent memory that uses energy landscapes instead of vector similarity. Retrieval is Boltzmann-weighted (temperature controls exploration vs exploitation), memories deepen with use (frequency = basin depth), and offline “sleep” consolidation merges, prunes, and strengthens — the same physics that governs CAPE coupling.

Why it’s different: Every existing agent memory (Mem0, Zep, Letta, MemGPT) uses semantic similarity + heuristics. Basin Memory uses a free-energy landscape — retrieval quality improves with use because the energy surface reshapes. It’s a learning theory, not a retrieval system.

Access: pip install basin-memory (pre-release). Request early access for the full physics engine.

Classifies superconductor pairing symmetry from bulk thermodynamics alone. 33/33 known materials classified correctly. Leave-one-out error 3.9%. Det(a) sign-change discriminator separates s-wave from sign-changing. L1/L0 boosting ratio fingerprints: 0.2x conventional, 3-5x s±, 11x d-wave.

Three falsifiable predictions from one framework: FeSe 8-pressure trajectory (β ~ 0.12), La3Ni2O7 ΔC/γTc = 1.7±0.3, UTe2 Leggett mode at 0.6–14 GHz (Method B ~4.6 GHz, unmeasured).

R² = 0.855, Bayes factor 1049.6. The same free-energy structure governs superconductor phase boundaries and AI scaling transitions.

Critical slowing down 2.60x, susceptibility 4.6x, Kramers escape rate within 1.4x of observation. Sleep consolidation connects to Basin Memory offline processing.