Agent-directed work package.
Target agent: Codex environment for SociOS-Linux/embeddinglab; Copilot coding agent if available.
Context:
embeddinglab is the Linux-native lab workspace for embeddings, rerankers, retrieval preparation, vector evaluation, and semantic indexing experiments for SocioProphet services. It should emit functional service manifests and evidence for governance; it must not self-promote models into production.
Scope:
- Inspect the live repository before editing.
- Add README explaining purpose and boundaries.
- Add
repo.maturity.yaml using repo-maturity.v1 semantics.
- Add
lab.manifest.json covering embeddings, rerankers, hybrid retrieval preparation, vector evaluation, and semantic indexing.
- Add
service-manifest/functional-service.v1.json conforming to SocioProphet/functional-model-surfaces semantics.
- Add
datasets/README.md, evals/README.md, adapters/README.md, training-runs/README.md, and examples/.
- Add Makefile validation.
Acceptance criteria:
make validate passes.
- Repo states lab-only boundary and governance handoff.
- Manifest can feed Holmes, Sherlock Search, model-router, and model-governance-ledger.
- No large model weights, datasets, secrets, or credentials are committed.
Boundary:
One PR only. Do not touch unrelated repos.
Agent-directed work package.
Target agent: Codex environment for
SociOS-Linux/embeddinglab; Copilot coding agent if available.Context:
embeddinglabis the Linux-native lab workspace for embeddings, rerankers, retrieval preparation, vector evaluation, and semantic indexing experiments for SocioProphet services. It should emit functional service manifests and evidence for governance; it must not self-promote models into production.Scope:
repo.maturity.yamlusingrepo-maturity.v1semantics.lab.manifest.jsoncovering embeddings, rerankers, hybrid retrieval preparation, vector evaluation, and semantic indexing.service-manifest/functional-service.v1.jsonconforming toSocioProphet/functional-model-surfacessemantics.datasets/README.md,evals/README.md,adapters/README.md,training-runs/README.md, andexamples/.Acceptance criteria:
make validatepasses.Boundary:
One PR only. Do not touch unrelated repos.