# Two batches, two actual model calls

On 24 October 2026, the packaged runner used Codex CLI 0.163.0-alpha.11.2 with GPT-6.1 Sol at High reasoning through an existing ChatGPT login. The [first report](first/report.html) covers 18 fictional feedback items. The [changed report](changed/report.html) covers a separate set of 12. Each run made one model call and completed without a retry.

Before either call, a separate input with a duplicate ID was rejected. Its recorded AI-call count was zero. Correct inputs then completed normally in fresh output folders. No source sample or earlier result was overwritten.

The model assigned themes and wrote observations; Python checked IDs and structure, calculated the counts and rendered the reports. Internal review checked all 30 items, source references, quotation flags, overlapping counts and ambiguous or injected text. It found no material semantic defect in these two outputs. That does not establish population accuracy, customer priorities or reliability on other material.

The second taxonomy was made from its own batch rather than copied from the first. It includes approval access and routing as separate themes. The two sets are not a panel of the same people and do not establish a trend.

Original classifications, assignments and input snapshots remain beside each report. [Execution identities](execution.json) bind the run to the packaged implementation. Public run summaries omit local command paths; original receipts remain in the private operating record. No account metadata or raw CLI logs are included here.

The execution receiver inspected the source data and generated HTML structure but did not open the reports in a permitted browser. Rendering, keyboard navigation and ordinary-user use are separate checks, not implied by successful execution. No schedule, automatic retry, public upload or extra cash purchase was part of these runs.