from-entangled-v2 continuous-v2

Status: mixed. The original results.md and analysis.md notes were drafted from a partial 33/36 snapshot, but the current local aggregate artifacts and the profiling pass now cover 36/36 completed runs.

This collection is the direct follow-up to from-entangled-v2/continuous-v1. It keeps the same six-task, six-loss-setting design and changes the primary modeling axis from continuous_latent_dim=8 to continuous_latent_dim=64. The current partial snapshot already shows that the higher-dimensional latent is not a uniform win: the ELBO anchor improves, the zero-beta anchor stays strong but slightly regresses, and the higher-dimensional low-beta and TWR regimes generally worsen router-side separation and reconstruction relative to the matched dim=8 reference.

A separate profiling pass now covers all 36 locally completed runs and is documented in profiling. That note focuses on profiled runtime, section-level time breakdowns, and the largest measured speedup opportunities.

• design
• results
• analysis
• profiling
• b0-const-deepdive-results
• b0p005-const-twr0p5-deepdive-results

Artifacts

• Analysis script: path:experiments/synthetic_sequences/disentangled/analysis/run_synthetic_pretrained_disentangled_from_entangled_v2_continuous_v2.py
• Aggregate directory: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v2-continuous-v2/
• Report: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v2-continuous-v2/report.md
• Findings: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v2-continuous-v2/findings.json
• Figures directory: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v2-continuous-v2/figures/
• Profiling run summary: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v2-continuous-v2/data/profiling_run_summary.csv
• Profiling sections: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v2-continuous-v2/data/profiling_sections_long.csv
• Profiling eval timings: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v2-continuous-v2/data/profiling_eval_long.csv
• Profiling speedup bounds: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v2-continuous-v2/data/profiling_speedup_bounds.csv

Context

• Direct design predecessor: from-entangled-v2 continuous-v1 design
• Direct results predecessor: from-entangled-v2 continuous-v1 results
• Prior layout references: synthetic-pretrained-disentangled-from-entangled-v2 results, synthetic-pretrained-disentangled-from-entangled-v3 results, synthetic-pretrained-disentangled-from-entangled-v4 results
• TWR caveat reference: synthetic-pretrained-disentangled-from-entangled-v5-twr design

Runs

• Slurmkit collection: synthetic-pretrained/disentanglement/from-entangled-v2/continuous-v2
• Planned design: 6 tasks x 6 loss settings x 1 seed = 36 jobs
• Current state: 33 completed and analyzable, 3 running
• Date range: 2026-04-21 to present