Value

D-Clean

Knowing the parameter is only part of the story.

Drag governs a simple dynamical system. Direct context and decoded physical parameters offer different ways to consume the same representation.

R² > .996still does not identify the best reader

Motion preview

Recorded trajectory · external force and drag · ½ speedProject-generated numerical trajectory replay. Source & changes

Across independent experiences

What persists.

Drag γ; mass is fixed at 1.

Across distinct realizations

What changes.

Initial state and piecewise force sequence.

Experience → prediction

The task.

Independent state/action history + recipient state.

→ Future state prediction.

Evaluated instance

The comparison.

JEPA · Align + Cross source. Null, Persistent, Decode and Oracle readers.

Controls. Adapted TDS / NOD / FCRL; physical-bottleneck readers.

Interpreting the evidence

What stays fixed in the test.

Persistent and Decode–Inject consume the same frozen donor code and share the frozen query encoder. Prediction horizon and consuming interface are controlled explicitly.

What this setting establishes.

Direct persistent context can outperform an accurate decoded-parameter bottleneck. In the reported h32 comparison, this occurs despite decoder R² above .996.

Scope of the evidence

D-Clean is an analytically integrated state environment with no native camera. Different baseline and downstream populations remain separate; these are prediction comparisons, not closed-loop control.

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 8

Adapted methods and native rollout references — C. D-Clean: common-reader H32

D-Clean report population; matched-reader H32 has 3 sources × 3 readers

Raw-state MSE ↓

Adapted methods and native rollout references — C. D-Clean: common-reader H32
Source	Matched	Fresh Null	Fixed Wrong	Fixed Zero
SPRII	0.00015072	0.01820554	0.05320123	0.03104658
State-space TDS	0.00134349	0.01820554	0.05257237	0.03143148
Released-component NOD	0.00618789	—	—	—
Paper-implemented FCRL	0.00133403	—	—	—

Bold marks the SPRII matched-reader error in this common-reader comparison.

SPRIIFixed Wrong and Fixed Zero change the fitted reader’s input; Fresh Null is trained separately.

Common-reader raw-state H32 errors use three sources and three readers. Fresh Null is fitted separately; Fixed Wrong and Fixed Zero change only the Matched reader’s input. Unevaluated interventions remain unavailable.

02 / Study Table 8

Adapted methods and native rollout references

D-Clean report population; matched-reader H32 has 3 sources × 3 readers

Raw-state MSE ↓

Adapted methods and native rollout references
Measurement	Source 0	Source 1	Source 2
Direct h16	0.00027058	0.00028337	0.00026174
Free rollout h32	0.00194772	0.00182717	0.00198696

Bold distinguishes direct prediction and free rollout; numerical errors are not pooled.

The dashed rule separates target horizon and rollout protocol.

Selected TDS source decoders are evaluated on the D-Clean report population. Direct h16 prediction and free rollout h32 use different prediction procedures and horizons; neither is the common-reader H32 comparison.

03 / Study Table 13

D-Clean source acquisition and donor routes

200 one-shot sealed-test systems; 3 source seeds

Drag probe R²; source-specific learned-embedding h16 errors

D-Clean source acquisition and donor routes
Method	Drag R^2 ↑	Self h16 ↓	Correct ↓	Gap ↑
Native	0.9932 ± 0.0007	0.0789 ± 0.0356	n.a.	n.a.
Structure	0.9935 ± 0.0007	0.0736 ± 0.0404	0.5512 ± 0.0681	0.0797 ± 0.0079
Same-rollout alignment	0.4434 ± 0.0367	0.0520 ± 0.0078	0.0542 ± 0.0090	0.00635 ± 0.00104
Align	0.9984 ± 0.0001	0.0140 ± 0.0040	0.0141 ± 0.0041	0.0948 ± 0.0037
Align + Cross	0.9986 ± 0.0002	0.0155 ± 0.0073	0.0155 ± 0.0072	0.0941 ± 0.0031
Align–Random	0.5299 ± 0.0400	0.0563 ± 0.0100	0.0591 ± 0.0100	0.00034 ± 0.00009
Supervised γ	0.9985 ± 0.0002	0.0119 ± 0.0024	0.0124 ± 0.0024	0.0969 ± 0.0044

Bold identifies strong drag accessibility under the two relation-trained alignment recipes.

The dashed rule separates the supervised-parameter reference.

AlignSelf and donor errors use each source’s own learned-embedding targets. They are not a common physical-unit leaderboard.

Means ± sample SD across three sources on 200 one-shot sealed-test systems. Self and donor errors use each source’s learned-embedding target. Correct uses another same-system rollout; Gap is shuffled minus correct. Native has no donor route.

04 / Study Table 28

Direct context versus decoded physical parameters — A. Absolute reader and intervention error

200 validation systems; 3 sources × 3 readers; same donor history; h32 primary

Raw-state MSE ↓; Decode-minus-Persistent gap; decoder R²

Direct context versus decoded physical parameters — A. Absolute reader and intervention error
Horizon	Null	Persistent	Decode	Oracle	Fixed Wrong
1	0.000577	0.000011	0.000013	0.000007	0.003227
2	0.001821	0.000020	0.000023	0.000009	0.009263
4	0.005689	0.000048	0.000056	0.000014	0.025559
8	0.015324	0.000120	0.000149	0.000031	0.061031
16	0.031972	0.000278	0.000349	0.000078	0.110287
32	0.059361	0.000590	0.000764	0.000223	0.180924

Bold compares direct and decoded context at the prespecified h32 endpoint.

32Oracle receives true parameters; Fixed Wrong is an input intervention on the Persistent reader, not a fifth fitted arm.

The four fitted readers use the same donor histories and frozen query encoder; h32 is primary. Fixed Wrong substitutes the Persistent reader’s context without fitting another reader. Oracle supplies true parameters to a learned reader, rather than an optimal-risk bound.

05 / Study Table 28

Direct context versus decoded physical parameters — B. Paired contrasts and intervention uncertainty

200 validation systems; 3 sources × 3 readers; same donor history; h32 primary

Raw-state MSE ↓; Decode-minus-Persistent gap; decoder R²

Direct context versus decoded physical parameters — B. Paired contrasts and intervention uncertainty
Horizon	Fixed Wrong 95% CI	D-P (10^-4)	D-P 95% CI (10^-4)
1	[0.002261, 0.004323]	0.0154	[0.0059, 0.0281]
2	[0.006822, 0.011988]	0.0291	[0.0076, 0.0586]
4	[0.019492, 0.032242]	0.0809	[0.0218, 0.1565]
8	[0.047516, 0.075592]	0.2898	[0.0965, 0.5504]
16	[0.086118, 0.136201]	0.7078	[0.2409, 1.3623]
32	[0.143277, 0.221349]	1.7394	[0.6824, 3.0741]

Bold marks the h32 Decode-minus-Persistent contrast and its paired interval.

32Positive D−P favors direct context. Contrast values use the stated 10⁻⁴ scale.

D−P is Decode minus Persistent; positive favors direct context. Gaps and intervals use the stated 10⁻⁴ scale. Intervals are pointwise paired-system 95% intervals, conditional on the fitted grid; h32 is the primary endpoint.

06 / Study Table 28

Direct context versus decoded physical parameters — C. Independent frozen-decoder validation

200 validation systems; 3 sources × 3 readers; same donor history; h32 primary

Raw-state MSE ↓; Decode-minus-Persistent gap; decoder R²

Direct context versus decoded physical parameters — C. Independent frozen-decoder validation
Sampling seed	Validation R^2	γ MSE	95% MSE interval
0	0.996715	0.003653	[0.0027, 0.0052]
1	0.997340	0.002957	[0.0022, 0.0043]
2	0.996993	0.003344	[0.0025, 0.0046]

Bold marks independently validated decoder R² for all three samplings.

0No sampling is selected as best. Accurate decoding does not determine the most useful downstream interface.

The frozen drag decoder is independently validated on 200 systems with 32 windows each, without fitting on validation data. Sampling seeds are shown separately. Accurate factor decoding does not establish which downstream interface minimizes task error.