Implementation sequence

September 21, 2026 · View on GitHub

Phase A — contracts + system_one_sdk 0.6.0

packages/system_one_contracts owns provider-neutral v1 request/response/model contracts, Choice ordering at the protocol seam, error/capability vocabulary, the inference provider behaviour and reusable provider conformance.

packages/system_one_sdk owns the rich semantic/client layer, bridges inference providers through SystemOneSDK.Providers.Contract, retains the hosted TypeSafe provider, and adds a generic Pristine-backed endpoint provider plus SDK-level conformance.

Phase B — dependency audit

Audit North-Shore-AI/crucible_safetensors, North-Shore-AI/hf_hub_ex, Bumblebee, Nx, Axon and EXLA. Determine whether crucible_safetensors 0.2.0 is actually required; verify immutable artifact/checksum/cache behavior in hf_hub_ex and ModernBERT/Qwen upstream requirements. Avoid private framework forks.

Phase C — system_one_bumblebee 0.1.0

Work in packages/system_one_bumblebee, in this order:

  1. Official/pinned Laya checkpoint and immutable artifact download.
  2. SHA-256 verification and safetensors parse.
  3. Inventory every checkpoint tensor; map every required tensor to Nx/Axon.
  4. Reproduce the exact tokenizer/preprocessor.
  5. ModernBERT encoder and Laya-specific decision head.
  6. Python oracle intermediate fixtures and CPU numerical parity.
  7. EXLA/GPU parity and resident Nx.Serving.
  8. SystemOneContracts.Conformance plus rich-SDK integration through the contract adapter.

This is real self-hosted execution with actual tensor weight import, not a Python-service wrapper. Later add Qwen option/logit scoring as a second architecture to prove the abstraction is not Laya-specific.

Phase D — system_one_server 0.1.0

Work in packages/system_one_server: GET /v1/models, POST /v1/systemone, health/readiness, auth, validation, telemetry, request IDs and resource limits. The server accepts SystemOneContracts.Provider implementations directly: first a fake provider, then native integration. Its request parser must feed the raw JSON body through SystemOneContracts.V1.Request.decode_json/1 so Choice order is preserved. Programmatic use does not require the server.

Phase E — lifecycle integration

External nshkrdotcom/self_hosted_inference_core: add :system_one_v1, launch/monitor/stop native services, attach to existing compatible services and return endpoint descriptors. Do not put inference execution in lifecycle management.