# Development disclosure

The original 24 case identities and corpus hash were frozen before their outputs. A first development pass used official Qwen3-0.6B Q8_0. On the first natural block the industry agent sometimes copied instruction text and treated clear technical problems as insufficient or routine service. That pass was stopped and preserved in `runs/development-qwen06-rejected/`; it is not the completed comparison and may contain partial blocks without receipts.

An intermediate official Qwen2.5-1.5B-Instruct Q4_K_M pass also gave poor semantic triage and overconservative criticism on the development demo. Its negative results are preserved in `runs/development-qwen15-rejected/`. The final model candidate uses the already-established public Qwen3-4B Q4_K_M revision and engine at the same source/ID/retrieval/candidate/inference ceilings. No hidden organizer labels were available. Natural cases N01–N06 were inspected during development and must be treated as development-exposed; N07–N18 were not inspected for that decision. Negative controls are constructed missing-evidence views, not hidden organizer no-match cases.

The earlier four-case development demo in `runs/development-initial/` also used Qwen3-0.6B, before prompt refinements and the strengthened two-document citation gate. It demonstrates an actual model/critic event path and records adverse outcomes, but cannot certify the final source version. Neural judgments and old experimental receipts must not be silently attributed to the new model.

These experiments do not establish a winning advantage or blind human usefulness score.