Research library

Claims you can trace back to the data.

Papers are presented as inspectable research records: question, preregistration, study design, results, limitations, provenance, and reproducibility materials.

SupportedNot supportedNot evaluable

01 · Featured study

Oracle Primitive Model

A sealed-test study of whether shared primitive organization improves generalization to withheld domain–operation combinations.

OPM v1.1.4Study closed · August 2026

Primary result

Shared primitives generalized across withheld combinations.

+42.62 pp

OPM recombination accuracy99.78%

Generalist baseline57.16%

95% bootstrap interval30.19–49.23 pp

H1 supported

Large recombination advantage with interpolation non-inferiority.

Mechanism not supported

The complete preregistered causal mechanism criterion did not pass.

20 primary runs · 4 conditions · 5 declared seeds · sealed evaluation

Read the research record

Publication standard

Readable as a narrative. Auditable as a dataset.

Claims first

Every conclusion is paired with its frozen threshold and formal status.

Data-shaped results

Tables use explicit conditions, splits, seeds, metrics, and uncertainty.

Immutable provenance

Authoritative artifacts are distinguished from pilots, diagnostics, and superseded runs.