synthetic-pretrained-disentangled-from-entangled-v2

Status: superseded - 17 analyzable cells, 4 collided local artifacts, 4 running cells, and 5 failed cells in the last local snapshot.

This collection is the first entangled-initialized loss sweep after v1, intended to test whether lowering or delaying KL pressure improves latent-strategy alignment. The local artifact collision issue prevents treating it as canonical, but the surviving cells still point toward lower effective KL pressure as the most promising direction and motivate the narrower v3 successor.

• design
• results
• analysis

Artifacts

• notebook: path:experiments/synthetic_sequences/disentangled/analysis/notebooks/13_apr15_ss_from_entangled_v2_analysis.ipynb
• aggregates: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v2
• exploratory figures: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v2/figures

Context

• depends on synthetic-pretrained-disentangled-from-entangled-v1
• superseded by synthetic-pretrained-disentangled-from-entangled-v3

Runs

• slurmkit collection: path:.slurmkit/collections/synthetic-pretrained-disentangled-from-entangled-v2.yaml
• run count: 17 analyzable completed cells, 4 collided completed cells, 4 running cells, 5 failed cells
• date range: 2026-04-15