This repo now includes a real native functioning model path that does not depend on unavailable external trainers.
A local exemplar-retrieval model trained from curated JSONL corpora already inside cognition-core and its supporting package ingestions.
- builds an artifact under
exports/candidates/<candidate>/exemplar_train/exemplar_model.json - loads that artifact through the
exemplaradapter - answers benchmark prompts by retrieving the strongest matching exemplars and synthesizing a bounded response
- can be benchmarked and scored end to end with the existing execution spine
- it is not a frontier generative model
- it is not LoRA / DPO / GGUF-derived yet
- it is a deterministic local specialist baseline proving that cognition-core can produce and run a real candidate today
It closes the biggest honesty gap: the repo now contains a working local model path instead of only scaffolding around future model paths.