ss-disentangled-direct-pretrained-pv-rmkl-sweep-v1

Status: complete - 21/24 cells complete and analyzable; 3 sorting_algorithms cells failed with CUDA OOM.

This collection is the first direct-pretrained follow-up sweep over posterior-variance and router-marginal-KL regularization. The main result is negative-but-informative: both regularizers mechanically control their target diagnostics, yet neither produces a reliable global improvement in strategy-aligned router latents.

• design
• results
• analysis

Artifacts

• notebook: path:experiments/synthetic_sequences/disentangled/analysis/notebooks/11_apr14_ss_direct_pretrained_pv_rmkl_sweep_v1_analysis.ipynb
• aggregates: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/ss-disentangled-direct-pretrained-pv-rmkl-sweep-v1
• exploratory figures: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/ss-disentangled-direct-pretrained-pv-rmkl-sweep-v1/figures

Context

• follows ss-disentangled-direct-pretrained-v1
• precedes the entangled-initialized alternative explored in synthetic-pretrained-disentangled-from-entangled-v1

Runs

• slurmkit collection: path:.slurmkit/collections/ss-disentangled-direct-pretrained-pv-rmkl-sweep-v1.yaml
• run count: 21/24 complete, 3/24 failed
• date range: 2026-04-14 to 2026-04-15