Evidence Layer

AI Lab

Engineering studies, validation protocols, and R&D experiments. Systematic analysis of AI architectures under real production constraints.

Internal coding-agent evaluation / engineering failure analysiscompleted

When Green Tests Lie: Coding Agents and Real Runtime Verification

Problem:

How should AI coding-agent output be evaluated and accepted when unit tests, static analysis, and generated reports all report 100% pass rates, yet a mission-critical user flow fails at the cross-runtime integration boundary under real execution?

Built / Tested:

The targeted engineering repair and verification architecture: serialization via URLSearchParams (application/x-www-form-urlencoded), exact Turnstile CSP allowlist, dual-tier rate limiting, strict allowlist for source_path, removal of dead TypeScript validation mirrors, and mock-free Playwright E2E testing against an actual spawned PHP runtime.

Internal AI/product architecture experimentactive

Perfume Engine: An Evidence-Backed Recommendation Prototype Without a Runtime LLM

Problem:

Determine whether a fragrance-alternative recommendation system can produce reasoned alternatives without collapsing source-backed facts, editorial assessments, system inferences, and uncertainty into a single opaque generative answer. The core question: can a recommendation experience remain genuinely useful while every public claim preserves an explicit provenance trail and deterministic inference boundary?

Built / Tested:

A controlled pipeline: private evidence → normalized claims → editorial assessment → deterministic inference → validated public projection → exact search → bounded alternative traversal → public explanation with full claim provenance.