Cross-Model Adaptation Portability

1. Mechanism and Primitive Description

The cross-model portability primitive operates on adaptations as portable artifacts. At authoring time, an adaptation declares the model context against which it was constructed: the model identifier, the model version, the input/output interface contract, and the assumed behavioral envelope (the range of model behaviors the adaptation relies on). It also declares portability metadata: which alternative models it could be ported to, which translation classes apply, and which behavioral assumptions are critical versus incidental. These declarations are part of the adaptation's claim and are signed into lineage at admission.

When the adaptation is to be transferred to a new base model, a portability translator is invoked. The translator is itself a credentialed authority within the mesh, with its own identity, claim (the source-model and target-model declarations it is authorized to translate between), evidence (the translation rules and validation procedures), and lineage. The translator consumes the source adaptation and produces a target adaptation: re-mapping interfaces, re-keying any model-specific tokens or embeddings, re-fitting any behavioral parameters, and validating the target's behavior against the source's behavioral envelope on a credentialed test suite.

The target adaptation enters the mesh as a new admission, but its lineage references the source adaptation, the translator's identity, the translation rules invoked, and any translation ambiguity surfaced during the process. Admissibility of the target is bounded by the model-version compatibility envelope: if the target model's version falls outside the envelope declared by the translator, the target is not admitted. If admitted, the target operates as a first-class adaptation, with the source-model lineage available for any audit, dispute, or further re-translation downstream.

The primitive supports lossy translation explicitly. A translation that drops or approximates some behavior records the lossiness in lineage so downstream consumers can decide whether to admit. It also supports fan-out: a single source adaptation may be translated to multiple target models, each producing an independent target adaptation with its own lineage thread.

2. Operating Parameters and Engineering Envelope

The primitive operates across a wide envelope of model-translation scenarios. Source-target pairs span minor version updates within a model family (e.g., perception-v3.1 to perception-v3.2), major version updates across architectural redesigns (e.g., transformer perception to state-space perception), and cross-vendor transfers (e.g., proprietary planner A to open planner B). The translation effort scales with the divergence: a minor update may be a re-key of token embeddings; a major redesign may require behavioral re-fitting; a cross-vendor transfer may require interface remediation in addition to behavioral adaptation.

Compatibility envelopes are parameterized by version range, behavioral coverage, and acceptable drift. A translator declares the version ranges it spans (e.g., source 3.0-3.5 to target 4.0-4.2), the behavioral coverage it preserves (e.g., 95% behavioral equivalence on the credentialed test suite), and the acceptable drift on out-of-distribution inputs. Adaptations that admit against this translator inherit those bounds; admissibility decisions downstream consult the envelope.

Validation requirements are also parameterized. Some adaptations require empirical equivalence on the credentialed test suite; others require formal property preservation (e.g., a safety-bound adaptation must preserve its safety bound); others tolerate approximate equivalence with explicit lossiness annotations. The translator's claim declares which validation class it provides; the source adaptation's claim declares which validation class it requires. Mismatched classes prevent admission.

Latency and computational cost vary widely. A pure interface-remap translation completes in milliseconds; a behavioral-fitting translation may require GPU-hours or longer of test-suite execution. The primitive supports both synchronous (admission blocked on translation) and asynchronous (translation pre-computed and admitted on demand) modes.

3. Alternative Embodiments

In an interface-remap embodiment, the translator only re-maps input/output interfaces and tokens; the underlying behavior is assumed to transfer because the target model is a near-equivalent. In a behavioral-fitting embodiment, the translator runs the source on a behavioral suite, captures the behavior, and fits the target to reproduce it within tolerance. In a distillation embodiment, the source serves as a teacher and the target adaptation is a student trained to match teacher outputs, with the credentialed test suite serving as quality gate.

In a multi-translator embodiment, several translators with different validation classes produce parallel target adaptations, each with its own admissibility envelope; downstream consumers select the target whose envelope matches their use. In a chained-translator embodiment, an adaptation is transferred A→B→C through composition of translators, with each step adding to lineage and the cumulative lossiness bounded by the chain.

In a self-translator embodiment, the source-model author publishes a translator alongside a model upgrade, allowing all adaptations against the prior version to be migrated automatically under the author's own credentialed authority. In a third-party translator embodiment, an independent credentialed entity publishes a translator and any adaptation can opt into the translation, useful for cross-vendor transfers and for situations where the source-model author has been deprecated.

4. Composition with Adjacent Primitives

Cross-model portability composes with the cascade-deactivation primitive: when a source adaptation is revoked, dependent translated targets are evaluated for cascade deactivation under the same dependency-graph rules; the lineage link from target back to source carries the cascade through the translation boundary. It composes with the credential-revocation primitive: a revoked translator invalidates targets it produced, and a re-credentialed alternative translator may produce replacement targets without re-authoring the source.

It composes with the lineage-audit primitive: an auditor inspecting a target adaptation can traverse back through the translator to the source, examine the translation rules and validation results, and verify that the admissibility envelope was honored. It composes with the dispute-resolution primitive: a contested translation (alleged behavioral divergence beyond declared envelope) enters a credentialed dispute path whose outcome may revise the envelope, recall the target, or recall the translator.

It composes with the byzantine-robust observer primitive when translations span untrusted vendors: multi-observer co-signature can be required for translations crossing trust boundaries. It composes with the marketplace primitive: translated adaptations enter the same admission and capacity-allocation flows as natively authored adaptations, indistinguishable to consumers except by lineage.

5. Prior-Art Distinctions

Conventional model-portability practice (e.g., ONNX export, framework conversion, knowledge distillation) treats portability as an engineering pipeline outside any governance frame. The translation is a developer-side activity producing a new artifact whose provenance to the source is informal at best. There is no credentialed translator authority, no declared compatibility envelope binding admissibility, no lineage that an auditor or consumer can traverse to verify the translation occurred under recognized rules.

Conventional skill-portability practice in robotics and agentic systems either rebuilds skills against each new model or hard-codes them against a single model with no portability. Where transfer learning is used, it produces a new artifact with no formal relationship to the source. Cross-vendor portability is rare and ad-hoc.

The present primitive differs in that the translator is a credentialed authority within the same governance frame as the adaptations it translates, the compatibility envelope is structurally bound to admissibility, the translation events enter lineage with full traceability, and the resulting target is a first-class adaptation whose provenance to the source is queryable. The five-property chain runs through every translation, providing audit, dispute, and cascade substrate that prior art lacks. Translation is a governance act, not an engineering side-channel.

6. Disclosure Scope

This disclosure encompasses cross-model portability of credentialed adaptations within the spatial mesh of provisional 64/049,409. The disclosure covers the portability-metadata schema declared at adaptation authoring (model context, version, interface contract, behavioral envelope, portability metadata), the credentialed translator authority and its claim structure (source-model, target-model, translation rules, validation class), the admission of translated targets bounded by the declared model-version compatibility envelope, and the lineage linkage between source, translator, and target. It covers the interface-remap, behavioral-fitting, distillation, multi-translator, chained-translator, self-translator, and third-party translator embodiments described above, and equivalent embodiments preserving the credentialed-translator and lineage-bound translation under the five-property chain.

The disclosure extends to the composition of cross-model portability with cascade deactivation, credential revocation, lineage audit, dispute resolution, byzantine-robust observation, and marketplace admission primitives of the broader mesh. It extends to alternative validation classes (empirical equivalence, formal property preservation, approximate-with-lossiness), alternative envelope parameterizations, and alternative translator-authority structures.

The disclosure does not depend on any specific model architecture, training framework, distillation technique, or interface format. It is technology-agnostic at the implementation layer and architectural at the primitive layer. Practitioners skilled in machine learning, model deployment, and credentialed governance will recognize the structural elements and may implement them using contemporary or future technologies provided the credentialed-translator authority, the model-version compatibility envelope, and the lineage-bound translation are preserved.