More AI document task: observed run and repairs, 7 October 2026

Configuration
More AI desktop 0.2.0 on macOS. The selected provider/model labels were
Z.ai (GLM) / glm-5.3-flash. File execution was on the computer. This does
not establish offline model inference: the inference endpoint was not
independently verified. No OpenCode comparison or separate Spanish run
was performed. The Spanish guide translates this English case.

Inputs and acceptance
The original CSV has eight synthetic tasks. Seven are done; T03 is pending
and its actual hours are unknown, not zero. Compare done tasks only:
planned 30, actual 33, variance actual minus planned +3 hours. The plan for
all eight tasks is 32 hours and is a different quantity.

The controlled revision changes only T05 actual_hours from 6 to 7. Expected
done totals become 30 / 34 / +4, and Boreal becomes 10 / 11 / +1. T03 must
remain unknown. Inputs and earlier output files were preserved.

What happened
1. More AI created an XLSX and a DOCX. During its own checks it repaired a
   blank-value guard that would have treated T03's missing actual as zero.
   Independent checks still found a material sign error: Tasks formulas
   used planned minus actual under an Actual minus Planned heading.
   Six nonzero row variances had the opposite sign; their sum was -3 while
   the correct summary was +3. The first XLSX was not accepted.
2. A specific follow-up asked More AI to correct the sign. Independent
   LibreOffice recalculation confirmed all eight rows, the unknown T03 and
   the 30 / 33 / +3 summary. The DOCX was unchanged and correct.
   Print review still found the last Tasks column on a separate page and
   cramped Summary headers. This version is not a print-quality example.
3. We supplied the one-value revised CSV and asked for new files. More AI
   produced the 30 / 34 / +4 result. Additional layout instructions and a
   separate completion request followed. The supplied prompts retain all
   these steps. No validator hand-edited the model output.

Verification
Independent read-only checks covered all source rows, task IDs, missing
values, formulas, project summaries, DOCX tables and prose, ZIP/XML package
structure and relationships. LibreOffice recalculated disposable copies;
the original model files were not replaced with these QA copies. Both
Word summaries were rendered and visually checked on one page each.
The revised XLSX was also exported and visually checked on two pages:
all seven Tasks columns share one page; Summary headers and the T03 note
are readable on the other. This check used LibreOffice PDF output.

The revised XLSX contains live formulas and requests recalculation on
opening. Its original numeric formula caches are absent: cache-only
previews can show blank results. Open it in a calculation engine before
checking or sharing it. Microsoft Excel/Word were not tested, and the
package checks are not full OOXML XSD validation.

Timing, usage and limits
The first two turns reported 1 minute 40 seconds and 37 seconds in the app.
Later turns were not independently timed. The app showed rounded usage
counts of 273.7k, 146.9k, 995.7k and 35.7k, and a rounded conversation total
of 1.5M tokens. These are UI observations, not a provider invoice or a
verified split of cached/input/output tokens. Monetary cost and cloud
credit consumption remain unknown; do not use these counts as a budget.

The automation accessibility tree truncated the long conversation. An
earlier assumption that a final answer was missing was withdrawn; no
product interruption is established by that truncated output.

Code execution was enabled only for the prepared synthetic test and
disabled in the app when background tasks were zero. The app also created
internal generator/backup files and memory.md in the test folder; the
instruction to write only in output was not an OS-level sandbox guarantee.

Archive contents
inputs/                 Preserved original CSV, notes and prompt template.
revised-input/          CSV with only T05 actual_hours changed to 7.
attempt-1-not-accepted/  First XLSX with the sign defect, plus original DOCX.
attempt-2-repaired/      Corrected arithmetic; earlier print issues remain.
revised-result/         Model files for the controlled input revision.
actual-prompts.json     Exact submitted prompts, including follow-ups.
qa-summary.json         Curated independent checks; not model output.
manifest.json           SHA-256 and sizes for the packaged evidence.

This is one synthetic task with repairs and follow-up checks. It is not a
first-attempt success claim, a general quality guarantee, or evidence of
superiority over another product. The deterministic expected files in the
other example downloads remain separate from these actual model results.
