A public archive of design artifacts and convergence logs from a running LLM agent verification engine.
The engine takes an abstract domain problem as input — "where does this design break?" — and runs an autonomous pipeline that generates, audits, rewrites, and converges until the semantic diff between spec and implementation reaches zero.
The core engine implementation is not public. This repository shows only: what was asked, what was found, and how it was designed.
Input (domain problem / observed failures)
↓
Designer — generates a design proposal
↓
Auditor — detects semantic drift between spec and implementation (DIFF)
↓
Writer — applies surgical corrections to resolve each DIFF
↓
[convergence loop: repeat until DIFF = 0]
↓
Reality Checker — probes whether the design actually works against real corpora
↓
Verifier — final validation
An anti-yes-man gate prevents the pipeline from converging on outputs that pass formal checks but carry no real meaning.
- Core engine, orchestrator, and templates are proprietary
- Only design artifacts and convergence logs are published
- For people who care about unconventional AI architecture
Feedback / Questions: GitHub Issues