What it solves
Decouple products from one model or hosted provider while keeping model identity, routing, failure behavior, evaluation, and release authority explicit and inspectable.
AI Plane · Accepted provider-v1 serving contour; portable customer productization remains in progress
AI Plane
Self-hosted model execution and control plane with a stable inference contract, replaceable models and runtimes, governed routing, provenance, evaluation, and evidence-backed release gates.
Good fit if...
Teams that need an owned inference boundary without coupling application code to one model, runtime, or hosted AI provider.
Not a fit if...
- Production-ready or universally customer-ready infrastructure
- Proven independent clean-host portability or dependency sovereignty
- Streaming, tool calls, or automatic hosted fallback in the current provider-v1 contract
- Universal support for arbitrary models, GPUs, or hardware profiles
Current status
Accepted provider-v1 serving contour; portable customer productization remains in progress
Public claims are limited to the capabilities documented below.
Implemented capabilities
- Stable provider-owned inference contract with health, readiness, logical model listing and authenticated non-streaming chat
- Logical model aliases backed by registry, manifests, lifecycle and provenance
- Local llama.cpp execution with an accepted Qwen2.5-Coder provider-v1 runtime contour
- Multi-candidate routing, bounded pre-dispatch fallback, circuit breaker and concurrency controls in source
- Evaluation, qualification and advisory release surfaces
- Clean-host preflight, non-mutating bootstrap planning and mandatory dependency inventory
Available within limits
- Provider-v1 serving is source-verified, target-runtime accepted and Human Release Authority accepted for an exact implementation
- Current main passes self-hosted source verification, but this does not extend runtime acceptance to all later Phase 2/3/4 source capabilities
- Self-hosted execution is materialized; independent clean-host portability and dependency sovereignty are not yet accepted
Limitations / not included
- Not production-ready, independently clean-host portable, dependency-sovereign or universally customer-ready infrastructure
- Streaming and tool calls are not supported in the current provider-v1 contract
- Hosted external fallback and automatic post-dispatch retry fallback are disabled
- Current accepted runtime is bounded to the declared llama.cpp/Qwen/dev-cpu contour, not arbitrary models, accelerators or hardware
- Not a mandatory inference provider for every Bennu module and not proof of consumer pairing
Delivery
- Bounded self-hosted provider-v1 runtime on a qualified controlled host
- Integration and design-partner evaluation against the stable inference contract
Planned / roadmap
- Transactional executable installer with logical target binding and independent clean-host acceptance
- Dependency-survivable acquisition, redistribution decisions and dependency-sovereignty acceptance
- Independent assurance and production-readiness acceptance
- Distributed and multi-node execution after portable single-host acceptance
Technical evidence
Proof & Evidence Status
Claim Snapshot: Verified (VERIFIED)
Current AI Plane README, Customer Value, inference contract, roadmap, and machine-readable release truth were reviewed; source implementation, accepted provider-v1 runtime, and incomplete Phase 5A productization are kept explicitly separate. (Reviewed: 2026-10-10)
Runtime / Operational Proof: Bounded (BOUNDED)
The provider-v1 serving contour has source verification, target-runtime acceptance, and Human Release Authority acceptance for exact implementation 510ac23f…; current main 225616b… separately passed self-hosted source verification with 1012 tests. This is not production readiness or independent clean-host acceptance.
Explicit Operational Boundaries:
- Runtime acceptance is scoped to the provider-v1 serving contour, not every Phase 2/3/4 source capability
- Independent clean-host portability is not accepted
- Dependency sovereignty is not accepted
- Independent assurance and production readiness are not accepted
- No universal support is claimed for arbitrary models, accelerators, or hardware profiles
Evidence is snapshot-scoped. Claim verification does not imply universal production readiness, certification, or SLA guarantees.
Read Proof Methodology & Evidence Model →