Move from starter spec pack to a benchmarkable cognition-core training program.
- Expand each seed dataset family to at least 50 examples.
- Expand each benchmark family to at least 20 tasks.
- Validate schemas on every JSONL and config file.
- Freeze the first real model trio in
configs/model_registry.yaml.
- Build repo-state capsule generator.
- Build benchmark runner with per-task result artifacts.
- Add minimal-prompt and bare-prompt benchmark profiles.
- Add quantization comparison report generation.
- Add training launcher stubs.
- Add checkpoint cards.
- Add candidate promotion reports.