Agent-native language models, built under evidence.
Ainfera builds Neptune — a clean-lineage program for agent-native models, a Stage-0 trust factory, and verifiable evaluation surfaces.
| Layer | Role |
|---|---|
| Neptune Core | Open proof layer: Neptune-1.0-27B and Neptune-1.0-9B |
| Stage-0 factory | Provenance, admission, training control plane, and release gates |
| Eval certificates | Preregistered held-out evidence and VAC/$ (Verified Agentic Completions per Dollar) |
| Closed specialists | Domain systems served privately; not open-weight by default |
Training completion is not a release criterion.
No Neptune model claims superiority, ships publicly, or supports external narrative without:
- Preregistered, held-out comparative evidence
- Provenance-bound reproducibility
- A passing release gate
- Website: ainfera.ai (in progress)
- Org profile: this repository
- Core model launches: announced with certificates when gates pass
Questions and partnerships: via ainfera.ai when live.