Measured result, bounded claim
On eight stratified local cases, structural validity moved 0.333→0.792 and executability 0.725→0.975; recall trade-offs and lexical limitations are published.
Flagship · Applied AI systems
v1.0.0 released · reproducible evidencePrompt Enhancer
A released local-first intent-to-specification compiler that treats model output as an unreliable dependency and makes provider behavior, fallback, resource use, and evaluation inspectable.
Engineering thesis
The hard problem is controlled orchestration.
Prompt quality alone is not a systems claim. The released system combines a typed provider contract, versioned protocol, authenticated remote exposure, bounded scheduling, deterministic fallback, capability-confined file operations, and correlated outcome traces.
The three-minute reviewer path is deterministic, offline, credential-free and independent of specialized hardware while exercising the same application boundary as local and configured providers.
Released evidence
What v1.0.0 proves
Provider-independent core
Deterministic mock, supervised local inference, and OpenAI-compatible adapters pass one typed contract and error taxonomy.
Failure paths are executable
53 Bun, 79 pytest, and 9 integration tests cover auth, bounds, cancellation, reconnect, supervisor failure, fallback, and path confinement.
Traceable artifact delivery
Eleven release assets include source, SBOM, dependency inventory, checksums, provenance, and raw benchmark evidence.
Evidence boundary
The release documents what it does not prove.
- The real-model comparison is an eight-case lexical evaluation, not human judgment or a general superiority claim.
- Local latency was measured on one RTX 3070 Ti workstation and is not presented as portable performance.
- Loopback limits network exposure but is not an operating-system sandbox; remote providers retain their own data policies.