Papers/OPM v1.1.4/Discussion

Scientific manuscript · Closed study

Oracle Primitive Model

Shared procedural computation for transfer to withheld domain–operation combinations.

Protocol
v1.1.4
Primary H1
Supported
Mechanism
Not supported
Study state
Closed · 27 Aug 2026
Chapter 08 of 09

Discussion, limitations, and implications

OPM v1.1.4 establishes a controlled behavioral transfer result, narrows several alternative explanations, and leaves the complete mechanism and efficiency program unresolved.

8.1 · Principal interpretation

Continuous primitive sharing changed what the model could do outside the training matrix.

Every condition learned the observed combinations to near-ceiling validation and interpolation accuracy. The decisive difference appeared on familiar operations placed in familiar domains but combined in a way training never showed. Under that intervention, OPM_SHARED remained near ceiling while the domain-generalist control and both untied procedural controls degraded substantially.

This pattern is consistent with the central transfer theory: an operation implemented by parameters shared across domains remains trained and available when requested in a held-out domain. Domain-specific parameters do not receive that cross-domain experience. Because DOMAIN_GENERALIST received shared operation embeddings and learned every observed operation within each domain, the contrast is stronger than comparing OPM only with an entirely untrained operation copy.

Established

Behavioral recombination advantage

A preregistered, sealed, seed-paired comparison supports H1 with a large interval-separated effect.

Partially supported

Primitive-path involvement

Large specific ablations and chance-level adapter-only performance suggest the primitive path carried necessary computation.

Not established

Complete causal reuse

The all-of criterion failed because one neural-probe family did not pass and two required inputs were unavailable.

Not tested

Practical efficiency and autonomous discovery

Resource, hardware, exception-boundary, and learned-decomposition hypotheses require new preregistrations.

8.2 · Alternative explanations

What the design rules out—and what it does not

The result is not readily explained by global training failure, because all conditions reached very high observed-cell validation and interpolation performance. It is not solely a comparison with an empty held-out module, because DOMAIN_GENERALIST had a trained transition for the target domain and a shared embedding for the target operation. It is also not evidence that oracle tokens directly encoded the label: the primitive path was restricted to selected facts and the adapter-only control returned performance to chance.

However, the experiment does not identify a unique internal ontology. Shared parameters may implement a computation that is functionally useful without corresponding one-to-one with a human-named relation. The study also cannot determine whether the same advantage survives learned parsing and routing, natural language, noisy bindings, broader task families, or domains with only partial semantic analogy.

8.3 · Frozen claim ledger

Every formal decision, including the negative ones.

6 supported2 not supported7 not evaluable
Claim identifierStatusObserved evidence / disposition
H1-PRIMARYSUPPORTEDΔgeneralist = 0.426200; 95% CI [0.301938, 0.492272]; interpolation checks passed
H2-THEORETICAL-RESOURCE-ADVANTAGENOT EVALUABLENo preregistered quality-matched resource threshold
H3-REAL-HARDWARE-ADVANTAGENOT EVALUABLEDeclared out of scope for v1
H4-CORRECT-BOUNDARY-OF-REUSENOT EVALUABLEDeclared out of scope for v1
MECHANISM-CAUSALLY-REUSABLE-PRIMITIVENOT SUPPORTEDComposite all-of criterion failed
MET-005-INTERCHANGE-MECHANISMNOT EVALUABLEAccuracy 0.346271; no preregistered decision threshold
MET-007-SURFACE-INVARIANCENOT EVALUABLEAccuracy 0.999354; no preregistered decision threshold
STA-004-ACTIVE-ABLATIONSUPPORTEDEvery active operation met the ≥0.20 drop requirement in ≥2 trained domains
STA-004-ADAPTER-ONLYSUPPORTEDMaximum split accuracy 0.500125 ≤ 0.60
STA-004-NEURAL-PROBESNOT SUPPORTEDOPM seed 4404, evidence step 1 failed Wilson/Holm criterion
STA-004-RAW-ORACLE-PROBESNOT EVALUABLERequired aggregate-v4 input unavailable
STA-004-RECOMBINATIONSUPPORTEDOPM_SHARED recombination accuracy 0.997811 > 0.80
STA-004-SENTINELSUPPORTEDMaximum logit change 0.0 < 1e-7 across 80 summaries
STA-004-SYMBOLIC-ORACLENOT EVALUABLERequired aggregate-v4 input unavailable
STA-004-UNRELATED-PRESERVATIONSUPPORTEDMean unrelated-operation drop was 0.0 for all four operations
8.4 · Limitations

The result is deliberately narrow.

  1. Oracle decomposition.

    Typed arguments, operation tokens, selected evidence, and execution order were supplied. The study did not test autonomous decomposition.

  2. Synthetic exact-isomorphism regime.

    The three rendered domains were constructed to share the relevant formal operations. Natural domains may share only partial or contested structure.

  3. Deferred baselines.

    Dense transformers, conventional MoE systems, oracle-informed MoE systems, and dense-execution primitive controls were outside the canonical v1 comparison.

  4. No quality-matched efficiency claim.

    H2 lacked a preregistered resource threshold; H3 hardware endpoints were out of scope.

  5. No domain-exception test.

    H4 requires a later partial-analogy regime with controlled exceptions and negative-transfer thresholds.

  6. Incomplete mechanism package.

    One neural-probe family failed, and symbolic-oracle plus raw oracle-probe inputs were unavailable in aggregate v4.

8.5 · Preregistered next studies

Each open theory becomes a separate experiment.

These identifiers are prospective. They do not amend or reopen the closed v1.1.4 result.

OPM-MECH-COMPLETE-001

Complete the mechanism evidence package and replicate neural probes.

OPM-MECH-INTERCHANGE-001

Preregister context-aware interchange controls and numerical thresholds.

OPM-RESOURCE-001

Test quality-matched parameter and FLOP advantages.

OPM-HARDWARE-001

Measure latency, throughput, memory, and energy on declared hardware.

OPM-BOUNDARY-001

Test domain-specific exceptions and negative-transfer boundaries.

Conclusion
Shared primitive organization improved controlled recombination generalization. The complete causal explanation remains an open scientific problem.