# One corrective chat follow-up

On 24 October 2026, the same GPT-6 Pro chat received one narrowly targeted corrective request. It was asked to remove superseded unavailable-cost claims, reconcile the companion files and retain the existing evidence limits. No new source facts, scenario or model were introduced.

[Original corrective PowerPoint](riverbend_saturday_decision_revised_v2.pptx) · [Original corrective Word companion](riverbend_outline_notes_source_map_revised_v2.docx)

The earlier two-turn result remains [a separate mixed trial](../receipt.md). Its four original files were preserved. These corrective files also remain unchanged.

## What improved

The six-slide PowerPoint and four-page Word companion retain Westbank's USD 3,720 staffing cost and the known combined pilot cost of USD 6,600. The old unavailable-cost statement and contradictory addendum were removed. Missing visits, the counter change, no calibration bridge and no expansion approval remain visible.

Both files opened, rendered and passed a representative editable-text save/reopen check on copies. All six slides and four Word pages were inspected without observed clipping, overlap or missing glyphs. This is a local rendered observation, not a timed room presentation or ordinary-user study.

## What still failed

The revised source-map entry gives the right new Westbank arithmetic, then retains a second Basis paragraph and cites the initial staffing_costs.csv and research_notes.md R3–R4. Those initial sources did not provide the new cost. It came from the later task update. The dedicated combined-cost source-map entry was also removed with the addendum.

This was a partial repair, not a clean provenance pass. No fourth model turn was made. Any separately labelled editorially reviewed reference is a later local repair, not an alteration of these original outputs or evidence that the chat corrected every source claim autonomously.