Abstract and research summary
A scientific account of the question, experiment, evidence, and limits of the Oracle Primitive Model v1.1.4 study.
Full manuscript
Study closed
27 August 2026
Can shared neural computation transfer procedures to combinations absent from training?
Neural systems often solve related tasks with representations whose degree of functional reuse is difficult to identify. The Oracle Primitive Model (OPM) isolates one part of that problem: whether shared procedural parameters improve transfer when the correct decomposition, typed argument bindings, and execution order are supplied. In a controlled synthetic environment, four neural conditions learned binary relational procedures across three rendered domains and four operation families. Training withheld three valid domain–operation combinations while retaining supervision for every domain and every operation elsewhere.
After a frozen 24-run pilot, the primary study trained four conditions across five declared seeds for 50,000 steps each. Checkpoints were selected using observed-cell validation only. The preregistered endpoint compared shared primitives with a capacity-advantaged domain-generalist control on sealed recombination data using a 10,000-replicate paired two-level bootstrap over seeds and latent worlds.
OPM_SHARED achieved 0.997811 recombination accuracy, compared with 0.571611 for DOMAIN_GENERALIST, producing a difference of 0.426200 (95% percentile interval 0.301938–0.492272). Interpolation non-inferiority checks passed, supporting H1. Renderer and structural tests also favored the shared condition. The complete causal-mechanism claim was not supported: active ablation, adapter-only, recombination, sentinel, and unrelated-preservation criteria passed, but one OPM seed-step family failed the preregistered neural-probe criterion, and symbolic-oracle and raw oracle-probe inputs were unavailable in the authoritative aggregate.
These results establish a behavioral phenomenon under oracle decomposition: continuous primitive sharing substantially improved transfer to withheld combinations without degrading interpolation. They do not establish theoretical resource advantage, real-hardware advantage, correct handling of domain-specific exceptions, autonomous decomposition, or the complete causal explanation. OPM v1.1.4 therefore supports the claim that shared procedural organization can improve controlled recombination generalization while leaving the precise mechanism and practical efficiency questions open.
compositional generalization · parameter sharing · modular neural networks · causal intervention · oracle decomposition · recombination transfer
One supported result, one unresolved explanation.
OPM_SHARED minus DOMAIN_GENERALIST on sealed recombination.
95% paired two-level bootstrap interval in percentage points.
Effect threshold and both interpolation checks passed.
One criterion failed and two required inputs were unavailable.
The scientific problem
A model may perform several related tasks without using the same internal computation to solve them. Similar outputs, shared labels, or correlated representations are not sufficient evidence of functional reuse. The central question is whether a network can apply the same learned procedure across domains and retain that procedure when a familiar operation appears in a domain where the combination was never trained.
OPM approaches this question as an upper-bound mechanism experiment. It supplies the correct procedure decomposition at first so that parsing, routing, and binding errors cannot obscure the narrower architectural test. If shared procedural computation fails under ideal decomposition, learning the decomposition would not rescue it. If it succeeds, later experiments can remove oracle support one component at a time.
Learn reusable computational primitives, compose them into procedures, and preserve only the domain-specific differences that causally matter.
The question tested in v1.1.4
Given multiple rendered domains that share formally specified operations, does continuous sharing of operation-specific primitive parameters improve accuracy on valid domain–operation combinations withheld from training, relative to a strong domain-generalist control and non-sharing procedural controls?
Across domains, within domains, or cloned and then untied.
All neural conditions received the same allowable bindings and procedure information.
Macro performance over three valid domain–operation cells absent from training.
Read the study as a paper, one chapter at a time.
The metric surfaces are now placed inside the chapter that explains their evidential role.
OPM works according to H1; v1.1.4 does not yet completely establish why it works.