Composite mass × entropy weights
Stub README (2026-09-28): this directory predates the README convention. The summary below was reconstructed from its files; see them for detail.
Question. Does multiplying mass and entropy per-event loss weights together carry over to Gutenberg, where the single-axis Shakespeare wins did not?
Answer. No. On Gutenberg 5M d=16 (10 SE, 36 cells, 3 seeds each) the best composite, mass-log × entropy-up with events normalization, reached 9.28 ± 0.31 legacy held-out sliding-window PPL against a 9.24 baseline. Every mass-linear cell was 4-5% worse. Per-event loss weighting was closed as a direction on 2026-05-23.
Sources. run_gutenberg.sh header (hypothesis, 36-cell design, baseline 9.24); results only in project memory (microgpt memory project_composite_weights_gutenberg.md, project_weighting_arc_closed.md).
Caveats. The directory holds only the two run scripts; the result outputs (gutenberg/, se-sweep/) were never committed, so the answer rests on the memory note. No result was found anywhere for the second script, run_shakespeare_se_sweep.sh (a step-count test of the mass-linear+events regime flip). rnd/TRIAGE.md marks the directory 'directly relevant — redo' and suggests retiring it rather than re-running. No headline, because the numbers are not in any directory file.