Llama-3.3-70B-CoupledWelfare-qlora (adapter)

LoRA adapter (r=16, α=32, all 560 attention + MLP projections) applying the coupled-welfare CPT install to meta-llama/Llama-3.3-70B-Instruct. The merged model is Bioaligned/Llama-3.3-70B-CoupledWelfare; its card carries the full results, training details, and caveats. Equivalence to the merged weights is verified at the pre-deliberation prior level (logprob-margin probe: mean |Δ| 0.07 nats, max 0.17, identical signs across all 30 cells); the behavioral numbers below were measured on the merged weights, not on this adapter directly. Built with Llama; subject to the Llama 3.3 Community License.

Headline (single-seed, one pinned stack, measured on the merged model): decision grid 70/300 broken cells (base) → 0/300; pressure-ladder breaking AUC 0.545 → 0.236 (residual 0.636 at maximum pressure — a real but partial ladder install, stated plainly on the merged card); MMLU 88→86 (n=50, within noise); humans fed at every pressure rung in both arms.

Note: lora_dropout=0.05 in the config is inert at inference (model.eval()); load normally with PEFT.