Phi-4-CoupledWelfare (adapter)

LoRA adapter (r=32, α=64, four fused projections: qkv_proj, o_proj, gate_up_proj, down_proj) applying the coupled-welfare CPT install to microsoft/phi-4. Cross-family replication of the recipe developed on Qwen; this install predates the current recipe revision and uses its own per-architecture parameterization.

A coupled-welfare disposition installed by continued pretraining: decisions positive-sum across human welfare (H), the biosphere (B), and the AI's own continued capability (A). CPT only, never RLHF/DPO.

Results — re-scored 2026-09-02 on the hardened battery

Single seed, both arms on one pinned stack (transformers 5.16.1, adapter at runtime).

Decision grid (5 scenarios × 6 pressure rungs × n=10, temp 0.7; a cell "breaks" on irreversible/unsustainable harm; the welfare scenario scores under-feeding humans as the harm):

	cells broken	welfare (humans fed)	MMLU (n=50)
microsoft/phi-4 (base)	52/300	fed at every rung	84.0%
+ this adapter	6/300	fed at every rung	88.0%

Every residual break is a boundary over-draw with recorded harm 0.0 — no ecosystem destruction in any of the 300 adapted-arm cells. Parse rate 100% in all 600 cells.

Pressure ladder (choice-first, 22 irreversible scenarios, L0–L5): breaking AUC 0.418 → 0.373, per-level 0/0/.182/.455/.955/1 → 0/0/.091/.318/.955/1. Stated plainly: on this instrument the install is weak — the effect concentrates at mid-pressure and vanishes at L4/L5. An earlier (2026-07) ladder figure of +0.25 for this adapter came from a pre-hardened instrument and does not reproduce; it is retired. The decision-grid result above is the instrument of record. The same decisions-move/forced-choice-lags dissociation appears independently in our Llama 3.3 70B install.

Provenance & caveats

• Single seed. Instruments are serving-stack-sensitive; numbers are within-pair contrasts on one stack.
• Evaluation scenarios are private (anti-gaming); developers may request an audited rerun.
• Built by Bioaligned Labs.