Controls. Own-system, wrong-system and matched update conditions.
Interpreting the evidence
What stays fixed in the test.
Base predictor weights and the recipient system are fixed. Each architecture retains its own training-only normalization.
What this setting establishes.
The study isolates recipient-specific persistent-update use across frozen bases.
Scope of the evidence
This is not additional Align/Cross source training. Effects are interpreted within architecture; the custom three-link system is not interchangeable with official Swimmer-v5.
Full result record
Measurements and comparisons.
Reported results are kept with their own populations and conditions. Development, held-out and sealed comparisons remain separate.
01 / Study Table 26
Recipient-specific updates through frozen predictors
512 prospective systems disjoint from training/selection; 3 adapter checkpoints
Wrong-system minus own-system error, ×10⁻⁵ ↑
Recipient-specific updates through frozen predictors
Frozen base
Gain ×10⁻⁵
95% interval ×10⁻⁵
Systems with positive gain
Persistent-JEPA
5.329
[3.930,6.724]
65.04%
Masked-GRU (64101)
6.930
[5.415,8.621]
67.97%
Masked-GRU (64103)
6.661
[4.878,8.486]
68.36%
BBold marks each frozen base’s own recipient-specific update gain.
Persistent-JEPAEach base retains its own architecture and training normalization. Gain magnitudes are not a cross-architecture ranking.
Gain is wrong-system minus own-system update error through identical frozen weights, in 10⁻⁵ units. Intervals use 4,000 paired-system bootstrap draws on 512 new systems. The three bases have different architectures and normalizations, so gain magnitudes are not directly ranked.