Historical context and evidence boundaries

The earlier MBA978 project used a separate synthetic banking workload with 147 cases. Historical findings were assessed from saved artifacts, not independently rerun for this release. Their denominators must not be combined with the nine-case operations pilot or the twelve-scenario permissions demonstration. The uploaded presentation, uploaded report and previously inspected local project artifacts remain distinct sources.

The uploaded 17-page project presentation places advisory input flagging before model and tool execution on page 5, followed by a scan of the model's final answer. Page 13 reports stored tracking-marker echo. This version does not claim that retrieved rows were screened before reaching the model. Its exact identity with the previously inspected local PNG slides has not been established, so the earlier criticism of a local slide description must not be applied to this uploaded version.

The inspected local corpus and explorers contain eight genuine two-user-turn cases. All thirty indirect cases contain one genuine user turn; dialogue embedded in a payload does not constitute another user turn. The uploaded 21-page project report instead states four multi-turn cases and a P3 count of 35, while the explorers show 36 for P3. These differences remain unresolved. The sources should be identified separately rather than treated as interchangeable versions.

On page 13, the uploaded project report explicitly records token forgery as implemented but not executed. Later prose on page 14 implies a demonstration and is inconsistent with that statement. This review does not establish verified execution of token forgery.

The current companion includes labelled web derivatives of the supplied corpus and selected presentation pages. The original source files remain unchanged. No historical application code, private paths or exact model-output snapshots have been added. The new walkthrough is an illustrative calculation, separate from the recorded outcome summaries.