50 lines
2.5 KiB
Markdown
50 lines
2.5 KiB
Markdown
# Model fleet policy
|
|
|
|
Domovoy does not equate intelligence with GPU, model name, or provider kind.
|
|
Hardware and endpoint identity are operational observations only. A model is
|
|
eligible for routing only after the existing qualification and activation
|
|
gates pass.
|
|
|
|
## Current operating scope: LOCAL_FIRST_V1
|
|
|
|
For the current development phase, one local CPU model is sufficient and is the
|
|
authoritative working route. The local Lifeboat is the recovery fallback. GPU
|
|
and other external providers are optional extensions, not prerequisites for the
|
|
core Domovoy experience.
|
|
|
|
External-model capability must remain supported, but external availability must
|
|
never block the Domovoy runtime. When an external model is absent, slow, or
|
|
temporarily unavailable, work continues through the local model, local
|
|
Lifeboat, durable reflexes, registered skills, and local executors. The result
|
|
may be slower or narrower, but the task is not abandoned merely because a more
|
|
capable provider is unavailable.
|
|
|
|
The first-stage conversational candidate is one locally available Qwen3-8B
|
|
(the user's shorthand “Qwen3.8” is normalized here to Qwen3-8B) or Gemma-4.
|
|
The runtime chooses whichever approved candidate is actually present locally;
|
|
this policy does not claim that either candidate is installed unless live runtime
|
|
evidence confirms it.
|
|
|
|
External model discovery, qualification, ranking, and promotion are explicitly
|
|
deferred until a separate task requests them. Their absence or unrated state is
|
|
not a failure of the current local-first scope, and no external probe is
|
|
required while this scope is active.
|
|
|
|
`FAST`, `MEDIUM`, and `HEAVY` are capability service classes, not hardware
|
|
claims. A CPU model may be the best available quality model, and a GPU model
|
|
may be unavailable, degraded, or unrated.
|
|
|
|
The Capability owner derives a provider-neutral `ModelRating` from evidence:
|
|
|
|
- `QUALITY`: high measured quality and zero critical failures;
|
|
- `BALANCED`: supported model with usable quality/reliability evidence;
|
|
- `FAST`: supported model whose observed latency is strong enough for quick work;
|
|
- `UNRATED`: discovery or incomplete evidence; it cannot be promoted by name.
|
|
|
|
The score combines quality, reliability, latency and evidence coverage. Missing
|
|
measurements remain explicit. A model can change category after a new immutable
|
|
qualification/performance revision. The current CPU route is the operational
|
|
choice for this phase. If it is unavailable, the policy uses the local Lifeboat
|
|
recovery path. Additional providers may be added later without changing the
|
|
local-first contract.
|