Bearing Context#

Bearing maintains evidence-backed model capability observations and compiles them into versioned catalogs. It answers what current evidence says a model can do under a disclosed execution and task scope. Datum answers which configured, available model should satisfy a role locally. Station and other runtimes own availability discovery and invocation.

Language#

Model Identity: A model id plus explicit revision and quantization. Revision and quantization may be null when unknown, but may not be inferred from the display id.

Execution Profile: The runtime, adapter, effective context, tool surface, hardware class, and workflow condition for one concrete runtime candidate.

Execution Scope: The applicability boundary carried by an observation. An exact scope requires complete profile equality and is mandatory for evaluation observations. A partial declaration scope matches only asserted dimensions; null means unknown/wildcard, while an empty tool surface means known-empty. Model-level declarations may use a global null execution scope.

Task Profile: The task family, suite/task identity, and evaluator that give an observation its behavioral scope.

Capability Observation: One immutable, provenance-bearing record about a model. An observation carries measurements, evidence, freshness, and uncertainty. Evaluation observations also carry an outcome and execution/task profiles.

Fact Measurement: A source assertion intended to hold for its canonical model, execution, and task scope, such as a declared maximum context. Different values for the same overlapping fact scope form a Conflict Set.

Sample Measurement: One measured result, such as a benchmark task outcome or latency. Different sample values are observations of variation, not conflicts.

Trusted Source Mapping: A reviewed exact association between one upstream row identity and its model, runtime, and workflow condition. Source adapters may report unknown rows but never infer this association from a display label. A derived observation content-addresses the exact mapping entry; the refresh policy, not the adapter, decides whether the mapping's review evidence is accepted.

Release Artifact Set: Multiple structured source files that jointly define one benchmark release. Each file keeps its own snapshot envelope and digest; the adapter accepts the set only when release identity and cross-file schema agree.

Acquisition Record: The exact release-wide snapshot references, URLs, byte digests, and fetch times returned beside imported observations. Acquisition authenticates where the artifact set came from. Observation evidence addresses only the task result, category assignment, and mapping that determine that observation, so unrelated artifact changes do not churn its identity.

Catalog Snapshot: A deterministic, content-addressed compilation of valid observations at a caller-supplied asOf. It retains observations and conflict sets and performs no model ranking by itself.

Runtime Inventory: A caller-owned list of models it can actually launch. Bearing may rank that list but never discovers or expands it. The list may be empty so callers can validate the same rank policy when none of their launchable candidates has a complete execution profile.

Advisory Projection: A caller-requested, evidence-bearing view of one measurement and aggregation for each runtime candidate. It preserves missing, stale, conflicting, and incomparable states and never affects ranking eligibility, scores, or order.

Approved Source Manifest: A reviewed, versioned list of model-intelligence sources and the exact origins, resolver adapters, artifact templates, bounds, freshness policy, attribution, and limitations within which discovery may run. Discovery can propose a new source revision inside that boundary; it cannot onboard a source or approve a model mapping.

Source Revision Proposal: A content-addressed result derived from exact resolver snapshot envelopes. It binds the approved manifest digest, resolver version, official entrypoint and derived-resource acquisitions, discovered revisions, and manifest-derived artifact locators. It is review input, not a capability observation or trust decision.

Product Boundary#

• Evals and external adapters produce observations.
• Forage/Lookout fetch, authenticate snapshot acquisition, and detect source drift. Bearing source adapters validate the handed-off binding but do not independently establish network origin.
• Traverse proposes fields from unstructured content.
• Survey carries candidate/review records.
• Surface derives trust status and freshness meaning.
• Flow owns scheduling and process transitions; Plumb monitors operational checks.
• Bearing normalizes model-domain observations and publishes catalogs/rankings.
• Datum combines catalogs with local providers, credentials, roles, and runtime inventory.
• Station discovers runtime availability, presents Auto routing, and invokes.