synthetic-pretrained-entangled-v1

Status: complete - 6/6 planned task runs analyzed, with 2 abandoned retry directories on base_conversion.

This collection is the pretrained-backed entangled source model for the later synthetic-sequences disentangling line. The headline result is that answer quality saturates across all six tasks while mixture quality remains task-dependent: multidigit_addition and base_conversion retain both strong strategy decodability and clean unique-strategy behavior, while ambiguity remains concentrated in grid_pathfinding and sorting_algorithms.

• design
• results
• analysis

Artifacts

• notebook: path:experiments/synthetic_sequences/entangled/analysis/notebooks/01_apr18_ss_entangled_pretrained_v1_analysis.ipynb
• aggregates: path:experiments/synthetic_sequences/entangled/analysis/aggregates/synthetic-pretrained-entangled-v1
• exploratory figures: path:experiments/synthetic_sequences/entangled/analysis/aggregates/synthetic-pretrained-entangled-v1/figures

Context

• collection role described in Synthetic Sequences Entangled Baseline
• contrasted downstream with ss-disentangled-direct-pretrained-v1
• used directly as the source model for synthetic-pretrained-disentangled-from-entangled-v1

Runs

• slurmkit collection: path:.slurmkit/collections/ss-entangled-pretrained-v1.yaml
• run count: 6/6 complete and included in quantitative analysis
• extra attempts: 2 abandoned base_conversion retry directories not used in aggregates
• date range: 2026-04-13