synthetic-pretrained-entangled-v1 Results

Collection State

From path:experiments/synthetic_sequences/entangled/analysis/aggregates/synthetic-pretrained-entangled-v1/data/task_status.csv:

planned tasks: 6
analyzable completed runs: 6
total abandoned retry directories: 2
abandoned retries are both earlier base_conversion attempts; the final completed base_conversion run is the one used in aggregates

Collection completeness and abandoned attempts by task. Caption: All six planned task jobs are complete and included. The only irregularity is base_conversion, which has two older abandoned retry directories in addition to the final successful run.

Final Snapshot

From path:experiments/synthetic_sequences/entangled/analysis/aggregates/synthetic-pretrained-entangled-v1/data/summary.csv:

mean sampled test final-answer accuracy: 1.0000
mean sampled test unique-strategy rate: 0.9275
mean sampled test ambiguous-strategy rate: 0.0723
mean final probe accuracy: 0.8167

Per-task final metrics:

Task	Final accuracy	Unique strategy	Ambiguous strategy	Probe mean	Best val loss
`list_summation`	`0.9999`	`0.9998`	`0.0001`	`0.7810`	`0.2510`
`sorting_algorithms`	`1.0000`	`0.8815`	`0.1176`	`0.7499`	`0.1714`
`grid_pathfinding`	`1.0000`	`0.6839`	`0.3161`	`0.7713`	`0.1169`
`linear_equation_solving`	`1.0000`	`1.0000`	`0.0000`	`0.6993`	`0.2268`
`base_conversion`	`1.0000`	`1.0000`	`0.0000`	`0.9326`	`0.1307`
`multidigit_addition`	`1.0000`	`1.0000`	`0.0000`	`0.9659`	`0.2316`

Final task-level metric overview. Caption: Final answer accuracy is saturated across the collection. The main variance axis is mixture quality and probe decodability, not answer correctness.

Strategy Mixture Quality

Mixture quality is not uniform across tasks.

grid_pathfinding has the highest ambiguous-strategy rate at 0.3161
sorting_algorithms is the second-most ambiguous task at 0.1176
list_summation is effectively unique at convergence, with only 0.0001 ambiguous mass
linear_equation_solving, base_conversion, and multidigit_addition are fully unique at the reported precision

Per-task final strategy occupancy:

grid_pathfinding: alternating=0.1740, down-first=0.3186, right-first=0.1913, ambiguous=0.3161
sorting_algorithms: bubble-sort=0.2970, insertion-sort=0.2948, merge-sort=0.2897, heap-sort=0.0000, selection-sort=0.0000, ambiguous=0.1176
base_conversion: decomposition=0.3180, repeated-division=0.2732, via-binary=0.4088
linear_equation_solving: divide-then-subtract=0.3273, inverse-ops=0.3679, subtract-then-divide=0.3048
multidigit_addition: left-to-right-partials=0.3483, right-to-left-carry=0.3041, rounding-decomposition=0.3476

Final sampled-test strategy distribution by task. Caption: Ambiguous mass is concentrated in grid_pathfinding and sorting_algorithms, while the remaining tasks show clean multi-strategy occupancy with no visible off-strategy mass.

Training Dynamics And Probe Structure

The task-level training curves show that answer accuracy becomes strong early, while mixture quality and probe trajectories remain task-dependent through the run.

Training dynamics by task. Caption: Validation loss, validation final-answer accuracy, validation unique-strategy rate, and probe mean evolve differently across tasks even though the final answer metric saturates everywhere.

The final position-wise probe summary also varies strongly by task.

highest final probe mean: multidigit_addition at 0.9659
second-highest final probe mean: base_conversion at 0.9326
lowest final probe mean: linear_equation_solving at 0.6993

Final probe accuracy by relative position. Caption: Hidden-state strategy decodability remains strong overall, but the final position-wise probe surface differs across tasks.