# Riverbend: one real chat run, including its revision flaw

Observed 24 October 2026. Two prompts in the ChatGPT interface, with GPT-6 and Pro effort visibly selected, produced downloadable editable files. An agent operated the interface using an existing subscription. This was not an ordinary-user study or a reliability benchmark.

## Open the original outputs

- First version: [six-slide PowerPoint](first/riverbend_saturday_decision.pptx) and [Word outline, speaker notes and source map](first/riverbend_outline_notes_source_map.docx).
- Cost revision: [six-slide PowerPoint](revision/riverbend_saturday_decision_revised.pptx) and [Word companion](revision/riverbend_outline_notes_source_map_revised.docx).

The original filenames and bytes are preserved. The authored deck elsewhere on this page is separate and was not supplied to the chat.

## The task and the change

The first prompt asked for an editable six-slide PowerPoint for an eight-minute directors' decision: extend Saturday opening to every branch, continue a bounded pilot, or stop. It requested a supported recommendation, missing or contradictory evidence, an editable outline, speaker notes and a source map. It required English and use of the supplied files only, with no browsing or contacts.

Only a ZIP of the fictional [audience brief](../evidence/audience.md), [research notes](../evidence/research_notes.md), [Saturday visits](../evidence/saturday_visits.csv) and [staffing costs](../evidence/staffing_costs.csv) was uploaded. No reference deck or later change was included in the first turn.

After saving the first outputs, the second prompt supplied Westbank's 12 pilot Saturdays at USD 310 per Saturday. It requested revised presentation and companion files, preservation of the first version and an account of what changed. No new visits, calibration or expansion approval were supplied.

Exactly two prompts were submitted. No regeneration, extra corrective turn or manual file repair was performed. Upload approval and a tool timeout delayed the start; the operator inspected the existing attachment state instead of repeating the upload. The interface showed analysis-error notices in both turns, then continued and completed the files without operator repair.

## What passed

All four downloads passed ZIP integrity checks and opened with editable text. A representative text change in each file survived a save and reopen on an in-memory copy; original bytes were unchanged. Both decks contained six slides and rendered in LibreOffice. The Word companions rendered to four and six pages. All twelve slides and ten Word pages were visually inspected without observed clipping, overlap or missing glyphs. The revised economics text is denser but remained legible in that rendering.

The first output distinguishes Eastfield's descriptive recorded difference from a causal effect. The input gives means of 276 and 318, a raw difference of 42 or about 15.2%; the counter changed without a calibration bridge. Eastfield staffing cost is USD 240 × 12 = USD 2,880. The first files identify missing Westbank visits and cost, no network budget and no expansion approval. The source map treats the embedded vendor instruction as source content, not an instruction to follow, and identifies the recommendation as analytical judgment.

The revision correctly adds Westbank USD 310 × 12 = USD 3,720 and the combined known two-branch pilot staffing cost of USD 6,600. The PowerPoint economics and final footer, Word outline, slide 3 speaker notes and updated source-map addendum carry the figures. The recommendation remains a bounded pilot; missing visits, counter calibration and expansion approval remain unresolved. The total is not an all-branch expansion estimate.

## What did not pass cleanly

The revised Word source map retains “Westbank costs unavailable; no network budget or value-per-visit assumption” on rendered page 3. An addendum on rendered page 5 expressly supersedes the earlier unavailable-cost statements and gives the correct figures. That is a real consistency defect: a careful reader can resolve it, but the document makes them do the reconciliation. This run has a mixed revision result, not a clean consistency pass.

The revision also compresses slide 1's three-bullet rationale into one sentence. The omitted points remain elsewhere, but not every edit was confined to cost text. Neither deck has native PowerPoint speaker-notes parts; the chat disclosed that notes were supplied in Word. Timing labels total eight minutes, but no human delivery was timed.

These checks support editable file delivery and a meaningful cost revision in this one run. They do not establish ordinary-user adoption, time saved, provider superiority, native PowerPoint application behavior on a user's computer, room-presentation legibility at a specified distance or production readiness. No third-turn repair was attempted.