synthetic-disentangled-Mar19

Status: complete - 62/63 runs completed; one sorting_algorithms + beta0p1_base run missing.

This collection is the sparse Mar19 synthetic sweep that first clearly separated high task accuracy from clean latent-strategy alignment. Its most important result is that token-weighted reconstruction changes the best achievable alignment on some tasks, while occupancy-oriented auxiliaries do not.

• design
• results
• analysis

Artifacts

• notebook: path:experiments/synthetic_sequences/disentangled/analysis/notebooks/03_mar19_analysis.ipynb
• aggregates: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-disentangled-Mar19
• exploratory figures: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-disentangled-Mar19/figures

Context

• early synthetic reference point for later collections in Disentangled Synthetic Sequences Reports
• followed by broader continuous-CVAE and discrete-intervention collections

Runs

• slurmkit collection: path:.slurmkit/collections/synthetic-disentangled-Mar19.yaml
• run count: 62/63 complete, 1/63 missing
• date range: 2026-03-19