Modelreg (Model Registry) — foundations RFC

Status

*ccepted*— 20260509. Sector bootstrap (skeleton + 5 impl tickets) landed as part of /k-go services/ai audit wave (Modo C). Q1 resolved: pragmatic HF Hub subset, not full mirror. Q2 resolved: koder-curated requires 2 curator approvals + passing benchmark threshold in eval/.

Summary

Catálogo Koder de modelos curadosavaliadosversionados — análogo HuggingFace HubMLflow Model Registry. Separado de `runtime` (que é executor).

Motivation

runtime/ puxa via Ollama mas sem catálogo: quais modelos a Koder valida? Quais retornam quanto em qual benchmark? Sem registry, cada produto escolhe modelo no escuro.

Scope

In

• Model entries (ID, version, license, weights URL, benchmarks)
• CRUD
• Tags (e.g., koder-curated, pt-br-friendly)
• Integração eval

Out (yet)

• Hosting de pesos (escopo runtime/kdb-blob)
• Serving (runtime)

Initial design

Surfaces

• backend/ — Go API + DB
• app/web — catálogo UI (futuro)

Key APIs

• POST /v1/models — registrar
• GET /v1/models?tag= — search
• GET /v1/models/{id}/benchmarks — eval results

Dependencies

• services/ai/eval — benchmarks
• services/ai/runtime — consumer
• infra/data/kdb-doc — storage

Relation to existing sectors

• Pré-requisito pra escolha informada de modelo no gateway/runtime
• Consome eval results

Selfhostedfirst analysis (5 gates)

• *1 Feature parity* zero
• *2 Performance* N/A
• *3 Stability* N/A
• *4 Capability* simples
• *5 Critical-path readiness* destrava decisão informada de modelo per produto

Open questions

• Q1: Reusar HF Hub schema ou próprio?
• Q2: Quem aprova koder-curated tag?

Next steps

1. Ratificar esta RFC (1 round de comments).
2. Criar sector dir services/ai/modelreg/ com koder.toml, README.md, skeleton.
3. Abrir tickets de implementação em services/ai/modelreg/backlog/pending/.
4. Registrar em meta/docs/stack/registries/self-hosted-pairs.md se substituir externo.