Choosing which models the vault runs

SkyKeep runs every AI step against a model runtime inside your own

deployment, and an administrator chooses which models. There are two

choices, not one, because the jobs have opposite shapes: reading a

document at upload happens once per document with nobody waiting, while

answering a question happens once per question with a person watching.

The vault will only offer models the runtime actually has, and it

labels every claim it makes about them — whether it characterised the

model itself, derived what it says from the size the runtime reported,

or measured it on this deployment. It will not invent a description of

a model nobody here has run.

Choosing a model chooses who writes the sentence. It never changes what

the sentence may be written from: your compartment grants decide that,

before any model is asked anything.

The ingestion model borrows the answering model, and says so

Given a vault with an answering model and no ingestion model of its own
When the vault resolves which model reads a document at upload
Then it uses the answering model and reports that it is borrowing it

A vault told which model to read with uses that one

Given a vault with a different model chosen for reading documents
When the vault resolves which model reads a document at upload
Then it uses the chosen model and reports nothing about borrowing

The embedding model never quietly borrows a chat model

Given a vault with an answering model and no embedding model
When the vault resolves which model turns documents into an index
Then it reports that none is configured rather than borrowing one

A model the vault has characterised says what it is good and bad at

Given a runtime offering a model this project has characterised
When an administrator looks at what the deployment can run
Then the model lists both strengths and weaknesses, marked as curated

A model nobody here has run is described only by its own numbers

Given a runtime offering a model this project has never characterised
When an administrator looks at what the deployment can run
Then the vault claims nothing about what it is suited for, and marks the description as derived

A model whose size the runtime never reported gets no description

Given a runtime offering a model reported with no size at all
When an administrator looks at what the deployment can run
Then the vault says it cannot tell anything about that model

A runtime that cannot be reached offers nothing rather than guessing

Given a model runtime that answers with something other than a model list
When an administrator looks at what the deployment can run
Then the vault offers no models and explains why

An embedding model chosen to answer questions is refused

Given a vault that has been asked to answer questions with an embedding model
When the vault checks what the choice would cost
Then it reports the choice as incompatible and names the model

A chunk longer than the embedding model can read is reported

Given a vault whose chunks are longer than its embedding model can read
When the vault checks what the choice would cost
Then it warns that most of each chunk is stored but unreachable by meaning

A coupling the vault cannot judge says so rather than passing

Given a vault using an embedding model this project has not characterised
When the vault checks what the choice would cost
Then it says it does not know, rather than reporting the choice as fine

A prompt is never silently truncated to fit the model

Given a question whose evidence is far longer than the model's context window
When the vault sends it to the model
Then the answer is marked as written from only part of the evidence

A prompt that fits is not marked as truncated

Given a question whose evidence fits the model's context window comfortably
When the vault sends it to the model
Then the answer is not marked as truncated