HF-GOLDEN-DUALITY

A smaller (512-dim, 6-layer) HibbieFormer with the same φ-ratio + temperature-duality regularizers trains more efficiently than the 143M V3.

Consolidated single-notebook variant of the Golden Ratio codebase, batch 2, seq 2048, lr 5e-4 — intended to run on the same 919M-token corpus in Colab.

• ◆Model scale (512×6 vs 768×12)

• →Zero training steps executed: in-kernel launch crashed on a required --corpus_file argument, and the bash launcher exited because the corpus was never uploaded to the session.

• ·Launcher referenced ephemeral Colab blob paths that no longer existed — the session was abandoned after two setup failures.

Not an experiment — a broken launch attempt, recorded as such. The registry keeps these because 'we never actually ran it' is a real and common outcome.