Agentic Engineering · LUN
Designed so generated code doesn't gain an identity without a separate verification pass.
LUN is a production factory that aims to turn every generated piece of code into an atom of knowledge, identified and versioned. Sealing a candidate is designed to require passing through a separate verification and judging process.
- Status
- Prototype
- Last verified
- 2026-07-18
- Access
- TODO_CONTENT_CONFIRMATION
The problem it addresses
Code generated by AI models mostly stays unidentified and re-producible; the same behavior gets rewritten in different places, and its correctness rests on the model's own say-so. LUN aims to reduce that uncertainty by tying generation to verification and registration steps.
How it works
You hand over a spec; the factory lays out the software's contract-bound skeleton.
A local model proposes a candidate for every gap.
The candidate goes through a check designed to include a verifier and a separate judging process.
A doubtful candidate goes into quarantine; nothing is sealed without a human decision.
Verified capabilities
Contract-bound skeleton generation
The factory lays out a contract-bound skeleton that openly announces itself when an empty body is called.
Fill-rate tracking
Progress is tracked numerically as a fill percentage over the skeleton.
Identified registration
Every sealed atom is designed to enter the registry with its own identity and revision history.
Architecture and trust model
Each claim below is labeled with where it currently stands — none of these are confirmed, final guarantees.
Production pipeline
- In development
Candidate generation uses local models; verification and judging are designed as a separate process.
- In development
Sealed atoms are planned to be stored in a registry under their own identity.
Independent judge principle
- Design principle
Keeping the model that writes the code physically separate from the judge that evaluates it is a design principle.
- Design principle
The principle adopted is that a doubtful candidate isn't auto-rejected — it enters a quarantine state left to a human decision.
Known limits
The judge and hidden-test flow aren't end-to-end verified
The independent judge process, hidden test cases, and registry/runtime flow are still in development; no public, independently verified end-to-end demo has been published yet.
Local model capacity is limited
Candidate generation relies on local models; fill rate and quality can vary on complex gaps.
Access and distribution shape not finalized
Usage terms haven't been confirmed by the project owner yet (see the status block).
Evidence
A public set of tests, demos, or sample output for this product is currently being developed — it will be added here once published.