CheckMyCoach is my primary research system, with human validation in preparation. InteractionKit is released, frozen methodological software, and Knowledge Compiler is supporting evidence infrastructure with partial provenance recovery. The remaining projects document formative evidence, infrastructure, or frozen prototypes; they are not presented as parallel validated contributions.
A bounded prototype that routes selected AI outputs, generates candidate revisions, and checks target removal and information retention. In a fixed 40-case development run, 15 cases were routed, 12 delivered outputs removed the predefined target feature, and none passed every target-and-retention check. I co-designed the evaluation decomposition, specified system behavior, and reviewed outputs against these requirements. Human validation is in preparation; human evaluation has not yet been conducted.
Released, frozen software for typed AI interaction experiments and contract checks. The release makes experimental structure inspectable; it does not establish construct validity or equivalence across implementations.
Artifact details →Supporting infrastructure for typed evidence objects, structural checks, and partial provenance recovery. Structural conformance does not establish source fidelity or downstream validity.
Artifact details →A failed execution-contract audit that produced zero interpretable scientific outputs. Retained as an engineering methods record.
Audit record →A synthetic LinUCB prototype. Its metrics describe simulated proxies, not human trust or intervention effects.
Simulation record →A small single-process runtime with append-only execution history, explicit report denominators, and fail-closed handling of ambiguous external execution. Interactive evidence is available; source and tests are published.
Interactive evidence →Portfolio rule: software conformance, constructed-corpus outcomes, simulation behavior, and human evidence are reported as different evidence classes. None is used as a substitute for another.