Skip to content

ADR-002 — External sources are admitted without authority transfer; records round-trip and remain portable

Status: Accepted, 2026-09-06; amended 2026-09-07 by accepted readiness decision D10; task/serving/outcome claim scopes amended through ADR-011 on 2026-09-08; registry, external-planner and evidence-host independence amended 2026-09-09 Affects: Target Architecture §1, §4, §5, §7, §8, §11; Framework Specification; Phase Plan Phase 0 exit gate, Phase 3 tiers Evidence: local NVIDIA AIConfigurator source at commit 77fd0773407b3683d8a671fe24a30a7110651b64; Hugging Face Hub documentation snapshot retrieved 2026-09-09; NVIDIA acquisition announcement and SEC Form 8-K dated 2026-09-03

Context

The competitive analysis claimed no continuous real-GPU evidence database existed. Verification found SemiAnalysis InferenceX (nightly sweeps across nine datacenter SKUs and six engines, version-annotated, full database released weekly), vllm-project/recipes (per-model YAML with vram_minimum_gb, min_vllm_version, verified hardware), and NVIDIA aiconfigurator (silicon-backed and modeled planning across TensorRT-LLM, vLLM and SGLang, with generated deployment targets and explicit accuracy limitations). Competing with funded upstreams on every covered datacenter-performance cell is not a solo-feasible Phase 3. Consuming and checking their output is.

NVIDIA's 2026-09-02 definitive agreement to acquire Hugging Face concentrates a major model registry and discovery surface, inference-provider router, hosted deployment surface, NVIDIA hardware/software planning, and the signals produced by those activities under one future owner. NVIDIA has publicly committed to keep Hugging Face open across models, frameworks, clouds, providers and compute platforms; that commitment is useful policy evidence but not a permanent technical boundary for this product. Hugging Face remains a first-class source and publication surface, while no registry, planner, provider, hardware vendor or host becomes the identity, recommendation or evidence authority.

Decision

  1. §8 may hold externally produced evidence. It enters at the authoritative-external level with its own freshness clock and provenance, and never masquerades as maintainer-lab or independently-reproduced evidence.
  2. Every supported GPU is an equal ExecutionTarget; no hardware class is categorically included in or excluded from maintainer execution. The scheduler selects work from an evidence gap, not from a fixed GPU matrix. It considers the required claim scope, missing/stale/contested evidence, user and model demand, prediction uncertainty or error, unresolved failures, freshness, expected cost, and whether adequate external evidence already exists.
  3. External evidence is preferred when it already satisfies the required claim scope and execution fingerprint. It is not duplicated without a recorded, measurable reason. Maintainer execution on H100 or any other datacenter GPU is normal when needed to establish an exact artifact/container boot, prediction delta, hardware-specific failure or remediation, accepted workload or quality result, or an explicit user-requested claim. The product still concentrates original evidence where gaps are largest; club-3090, InferenceX and vLLM Recipes remain evidence sources at their declared levels.
  4. deployment-plan.json must round-trip, losslessly for the shared fields, to three external formats: a vllm-project/recipes YAML entry, an aiconfigurator estimate request, and an InferenceX-shaped result row. This is a Phase 0 exit-gate criterion.
  5. Signed community reports (§13) use the recipes-compatible shape so that a verified community configuration can be upstreamed with one action.
  6. Public data license: CDLA-Permissive-2.0 for evidence records, CC-BY-4.0 for documentation, chosen before record #1 is published. The canonical publication unit is a portable, content-addressed evidence release with a manifest, record digests, schema version and signatures/attestations where applicable. Record and release identity does not contain a host URL. Hugging Face Hub is the initial public dataset mirror, not the sole store or authority; the same release can be reconstructed from its archive and published through another conforming publisher without changing record ids.
  7. Every produced or ingested record carries a typed claim_scope: config, boot, memory, remediation, serving_performance, task_outcome, or outcome_economics; a production_mode; and a machine-readable reason. Historical performance and quality values migrate losslessly to the more precise scopes with their original value preserved. A boot record cannot become serving or task evidence; a serving benchmark cannot become task evidence; and a task result for a different task/application/evaluator/solution fingerprint cannot qualify the requested solution (ADR-011).
  8. Artifact location and artifact identity are separate. An ArtifactLocator names a source kind, source URI and requested revision. A conforming ArtifactSource resolves it to a source observation containing the resolved immutable source revision, manifest, per-file/component content digests, sizes, license/gating observations and attributed publisher metadata. ArtifactIdentity is derived from the resolved contents and component structure, not from a Hugging Face repository id or any other registry namespace. Equal resolved artifacts retain their distinct locations and provenance but compare as the same content identity; different contents never collapse because their names match.
  9. Publisher metadata is an attributed claim. Model-card fields, pipeline_tag, declared lineage, license labels, evaluation results and provider mappings remain artifact_declared or source observations until the applicable deterministic resolution or execution evidence promotes a narrower claim. A registry UI, security scan or provider widget cannot establish engine support, endpoint exposure, task fitness or deployment qualification.
  10. Add an external Planning source contract. AIConfigurator, provider recommenders, recipe planners and future systems may return a PlanningClaim containing producer and version, exact input fingerprint, proposed configuration, claim scope, epistemic status, source-record references, calibration domain, uncertainty and unsupported/opaque fields. A planning source proposes candidates and priors only. It cannot label a solution Recommended, Qualified, best observed or measured-efficient; the common deterministic qualification and selection policy owns those promotions. The owned mechanism-aware calculator remains the transparent GPU-free core and fallback rather than being replaced by an external planner.
  11. Evidence-gap scheduling uses independently attributed signals. Accepted user requests and local failures are primary direct demand; opt-in aggregate usage, multiple registries, provider catalogues, engine issues/releases, external planners/recipes and dated popularity/trend feeds may contribute. No single registry metric defines eligibility or priority. The scheduler records the source snapshot, normalization, deduplication and material contribution of every signal; popularity cannot override claim scope, uncertainty, failure value, adequate existing coverage or accepted policy.

Consequences

  • Phase 3 cost is controlled by evidence-gap scheduling, reuse of adequate external records, and the backend spend caps; it is not controlled by excluding a GPU class.
  • The tool's own records can be cited by, and can correct, the three external sources; this is the distribution mechanism.
  • Any external source that stalls simply ages out through its freshness clock; the tool does not depend on any one of them being alive.
  • A registry outage or ownership change can reduce discovery coverage but cannot change an already resolved artifact identity, erase a portable evidence release, promote a planner claim or alter a prior verdict.
  • Hugging Face remains the initial high-value artifact and dataset adapter. Registry neutrality does not require mirroring model weights, duplicating every registry, or withholding useful integrations; it requires source-neutral permanent contracts and tested failure behavior.
  • Amendment to §7 wording: "provision or attach hardware" now explicitly includes "or attach externally produced evidence at its declared level".