Zero-day model onboarding: compile release facts before support code

Observed: 2026-08-27T15:35:00Z Issue: #9421 Study receipt: study_a4310abb38820a69d2b77ffaa1ada12058d1cde24169bf651806ef1c9baa8d01 Current authority: Adding a new model to fak

This is the provenance note for the release-to-descriptor compiler. It records the dated external evidence and the design choice; the playbook carries current operating instructions.

Source ledger

Source State and event Immutable anchors read Platform context Refresh trigger
vllm-project/vllm shipped at 76c0c6530e832c758f5502ca4ef37ae88a7ffa13, 2026-08-27T15:30:22Z docs/contributing/model/README.md:3@76c0c6530e832c758f5502ca4ef37ae88a7ffa13; docs/contributing/model/registration.md:14@76c0c6530e832c758f5502ca4ef37ae88a7ffa13; docs/contributing/model/tests.md:10-27@76c0c6530e832c758f5502ca4ef37ae88a7ffa13; vllm/model_executor/models/registry.py:913-998,1183-1295@76c0c6530e832c758f5502ca4ef37ae88a7ffa13; tests/models/test_registry.py:31,206@76c0c6530e832c758f5502ca4ef37ae88a7ffa13 native model registry plus Transformers compatibility backend vLLM registry/compatibility contract change or a relevant revert follow-up
sgl-project/sglang shipped at 20a491d1d311553bbab3f22e19bbafb86ef3c0cc, observed 2026-08-27 python/sglang/srt/models/registry.py:61-134@20a491d1d311553bbab3f22e19bbafb86ef3c0cc; docs/docs/supported-models/support_new_models.mdx:8-105,315-433,515-517@20a491d1d311553bbab3f22e19bbafb86ef3c0cc SRT architecture registry, processors, tests, docs, optional dependencies, and generic Transformers fallback SGLang model-registration or fallback-contract change
vLLM history shipped then reverted/migrated, 2026-08-21 through 2026-08-24 native remove/revert/migrate: d53b1c2efc0ca7161c3844f51ca9acbcdb7129d5, e3f60265032e7c1e158a749d82a8c4dad62f941a, 48d7132962ef3e2e6982a316013294c0029e3a64; fusion/revert: 463aa5e30fe0b51a5d6ef22897962fc1be97cb85, 592e06f2ae115cbb0f7e2e1e776c255a3fe6c3c1 Hunyuan native/Transformers migration and Kimi-K3 ROCm KDA-prefill fusion the next native-model removal, backend migration, or architecture-fusion revert

Both roots were Apache-2.0 at the pinned revisions. The disposition is INSPIRE-ONLY: this spine copies no expressive upstream code, tests, or comments. Any later direct port requires its own per-file provenance and notice review.

What the sources prove

vLLM and SGLang optimize for breadth and contributor throughput. A newly released Hugging Face architecture can often run through a compatibility backend before a native implementation exists. Native support then fans out across architecture registries, model/config predicates, processors, weight loaders, tests, docs, dependency gates, and backend-specific behavior.

That is a useful zero-day ingestion strategy, but it is not a durable native-support contract. The vLLM history above shows two failure patterns that registration alone cannot prevent:

  1. native-model removal, revert, and re-migration can happen within days; and
  2. an architecture-specific fusion and its device witness can be added and reverted together within hours, erasing the runnable evidence along with the implementation.

FAK should borrow the speed of metadata intake without borrowing automatic foreign execution. A pinned release must become one deterministic packet before any code generation, allocation, or support claim. The packet keeps provenance, semantic uncertainty, native/backend/test/oracle/docs/ performance work, registration closure, coupling cost, and rollback obligations together.

Inward witness and seam

The durable study recorded an ABSENT-on-axis verdict:

Issue #9421 therefore connects those existing seams rather than adding another registry, planner, or receipt stack.

Candidate disposition

Borrow Axis Source fact FAK decision Disconfirming check
Compatibility-first release intake time to ingest a newly published architecture both frameworks can recognize or route models before native specialization is complete DEFAULT / INSPIRE: accept pinned facts immediately, but compile only a fak-native work packet reject if three unrelated releases cannot be expressed without executable hooks
One release descriptor registration closure architecture, fixture, defaults, dependency, and docs state is spread across handwritten surfaces DEFAULT / INSPIRE: normalize one manifest through modeldescriptor and emit exact open obligations reject if the compiler hides model-specific work or creates a second registry
Typed semantic refusal false support avoidance generic compatibility fallback can make unknown architecture deltas appear runnable DEFAULT / DIVERGENT: unknown or contradictory deltas stop before allocation; no automatic fallback reject if a refusal can still emit scaffold/executable behavior
Persistent rollback witness learning across reversions rapid native and fusion reverts can remove their only witness FOLLOW-ON: keep oracle/performance/rollback obligations independent from an optimization patch reject if retained witnesses cannot reproduce the reverted failure

Completeness and honest boundary

The underlying deep study covered registry/compatibility code, model-contributor docs, registry and initialization tests, architecture-specific runtime mutation surfaces, recent reverts and repair history, and root/per-file license/provenance. It did not establish that every current vLLM or SGLang model is expressible by the FAK manifest; that remains the disconfirming replay for the follow-on agentic delta extractor. This note makes no runtime, quality, kernel-fusion, or performance claim.

Companions: field-borrow | study-repo | #8011 | #9282 | #9421