Authoritative sealed-test results
The primary behavioral hypothesis was supported: shared primitives produced a large, seed-consistent recombination advantage while preserving interpolation performance.
Primary finding
A strong behavioral result.
An incomplete causal explanation.
OPM_SHARED preserved near-ceiling performance on combinations withheld from training while the principal domain-generalist control fell to 57.16%. This supports the preregistered behavioral hypothesis, but it does not by itself establish the complete mechanism claim.
Δgeneralist on the sealed recombination test · 10,000 paired two-level bootstrap replicates · PCG64DXSM seed 99117
What the primary estimate means
The primary estimand was the paired difference between OPM_SHARED and DOMAIN_GENERALIST on the three withheld domain–operation cells. Model seeds and latent worlds were resampled in a two-level bootstrap. The observed difference was 0.426200, and the entire 95% interval remained above the preregistered +0.02 support threshold.
The interpolation non-inferiority conditions also passed. OPM_SHARED achieved 0.999347 on observed combinations, compared with 0.997264 for DOMAIN_GENERALIST and 0.997583 for PROC_UNTIED. The result therefore cannot be explained as trading ordinary interpolation competence for recombination performance.
5.2 · Authoritative outcomes
Sealed-test accuracy by condition.
Values pool the five declared primary seeds. Recombination is the preregistered H1 endpoint; interpolation, renderer, and structural tests provide context.
| Condition | Interpolation | Recombination | Renderer | Structural |
|---|---|---|---|---|
OPM shared primitivesOPM_SHARED | 0.999347 | 0.997811 | 0.999354 | 0.998979 |
Domain generalistDOMAIN_GENERALIST | 0.997264 | 0.571611 | 0.590729 | 0.890854 |
Procedure untiedPROC_UNTIED | 0.997583 | 0.499989 | 0.738417 | 0.873667 |
Procedure clonePROC_CLONE | 0.997861 | 0.500000 | 0.749292 | 0.873854 |
One model profile across all four sealed tests.
Interpolation was nearly saturated for every condition. Separation emerged when rules had to transfer to withheld combinations and altered renderings.
Interpolation
Recombination
Renderer
Structural
Recombination performance for every primary run.
Each mark is one frozen model seed. OPM_SHARED stayed between 99.17% and 99.97%, and every paired seed comparison favored OPM_SHARED.
OPM_SHAREDDOMAIN_GENERALISTPROC_UNTIEDPROC_CLONEThe smallest paired advantage was 17.91 points at seed 4404; the pooled lower-bound support threshold was 2 points.
Expected calibration error exposes baseline overconfidence.
Lower is better. OPM_SHARED remained below 0.002 on every split; non-sharing baselines were sharply miscalibrated on recombination and renderer tests.
5.3 · Result explorer
Interrogate the sealed metric cube.
Choose a model condition, test split, and analysis dimension. The explorer reads directly from the frozen baseline-metrics package.
Loading the metric cube…
Explore all 168 frozen dimension-level accuracy rows. Download dimension-accuracy.csv for analysis.