0026. A specialized local model is an optional Suggester optimisation
- Date: 2026-09-02
- Status: accepted
Context
Small open-weight models are unreliable at choosing the right Catalog operation, honouring Intent qualifiers and producing the Proposal schema. Fine-tuning one for Loomarr could help, but it could also become a second authority or a hard dependency of local AI.
Decision
A Loomarr-tuned model is an optional optimisation of the existing Suggester, never a new authority and never a prerequisite. It learns tool choice, qualifiers, schema and abstention, not title identity, ratings, quotas or authorization: every pick still passes the grounding chokepoint (0004). Certification comes first and may end the project; hard gates are pass/fail, not terms in a weighted score.
Training and conversion live in the separate loomarr-models repository. Loomarr consumes a
certified model only through its HTTP provider boundary; no trainer, converter or weights enter
this repository or its image.
Consequences
- The ordinary AI profile stays model-less; hosted and custom providers stay first-class.
- A new model version is pulled explicitly, verified by digest, probed and canaried before activation, and the previous one stays available for rollback.
- The experiment protocol is archived in
local-model-experiment.md.