Seal Balance/Homeostasis and Ghost-AuxK as two rejected research paths. Adds 42 verified runs without duplicating the nine already-public parent controls, plus compact/full browser payloads, reproducibility snapshots, and the updated root dataset card.
Seal Balance/Homeostasis and Ghost-AuxK as two rejected research paths. Adds 42 verified runs without duplicating the nine already-public parent controls, plus compact/full browser payloads, reproducibility snapshots, and the updated root dataset card.
Five matched 125M MoE runs: baseline plus gradient/update load scaling at c=0.5 and c=1. Includes full raw logs, provenance ledger, compact snapshot, full report payload, and reproduction card.
Describe the four-panel mean-plus-or-minus-one-SD and SD views, and record why expert-indexed load curves are excluded from cross-seed aggregation. Raw logs and snapshots are unchanged.
Clarify that raw riglogs remain full-resolution while compact snapshots may retain selected non-loss curves; describe the GitHub report catalog relationship accurately.
Point all report and study-browser links to their tracked docs/reports locations on GitHub; Hugging Face retains only raw logs, reproduction information, and compact snapshots.
Point the seed-variance report to docs/reports on GitHub and clarify that Hugging Face retains raw logs, reproduction information, and compact curve snapshots rather than report HTML or bespoke plotting scripts.
Add 41 clean 60M seeds and 22 125M seeds with full rig logs, provenance, reproduction notes, and compact curve snapshots. Seed 1369 is excluded from the 60M cohort because it used a different dirty recipe source.
Add 41 clean 60M seeds and 22 125M seeds with full rig logs, provenance, reproduction notes, and compact curve snapshots. Seed 1369 is excluded from the 60M cohort because it used a different dirty recipe source.
