exp2_discrete_cvae-Apr3-7loss_20260403

Status: complete - 144 completed runs analyzed from the extended Apr 3 / Apr 7 grid.

This collection is the discrete-CVAE intervention study on normalization, token-weighted reconstruction, and JSD. It turns the earlier broad diagnostics into matched contrasts and task-specific winner analysis. The strongest stable result is that reconstruction normalization helps; the more ambitious TWR and JSD claims remain more conditional.

• design
• results
• analysis

Artifacts

• notebook: path:experiments/synthetic_sequences/disentangled/analysis/notebooks/08_apr4_discrete_cvae_7loss_analysis.ipynb
• aggregates: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/exp2_discrete_cvae-Apr3-7loss_20260407
• exploratory figures: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/exp2_discrete_cvae-Apr3-7loss_20260407/figures

Context

• builds on the continuous-CVAE diagnostics in continuous_cvae-Apr1-7loss_20260401
• continues the disentangled synthetic sweep line from synthetic-disentangled-Mar19

Runs

• slurmkit collection: path:.slurmkit/collections/exp2_discrete_cvae-Apr3-7loss_20260403.yaml
• run count: 144 completed runs in the analyzed extended dataset
• date range: 2026-04-03 to 2026-04-07