## TL;DR
The agent filled gaps in a truncated file with plausible numbers. Never let it invent data URIs validate the export is complete before counting, re-fetch when it is not, and mark unfetchable values as unknown, not as guesses.

```text
agent hallucinated seat counts from a truncated CSV, audit failed
```

## Use this when
- Seat counts look suspiciously round or convenient.
- The source CSV was truncated.
- The agent produced numbers without a complete input.

## Not for this skill when
- The export is complete and counts are wrong. That is a logic problem.
- The numbers are estimates by design. Then label them as estimates.
- The agent refused to answer. That is over-caution, a different tuning.

## Steps
1. Check the export for truncation: abrupt end, row count versus expected. Verify: completeness is assessed.
2. If truncated, re-fetch the export; do not proceed on partial data. Verify: the new export is whole.
3. Mark values that cannot be fetched as unknown in the output. Verify: unknowns are labeled.
4. Add a completeness check before any counting step. Verify: the check gates the count.
5. Confirm every reported number traces to a data row. Verify: the audit is grounded.

## Variant phrasings
### hallucinated counts
The short phrasing.
### truncated CSV
The cause-first search.
### agent invented data
The behavior-first phrasing.

Compatibility: Any agent producing counts from exported files.

## Why it happens
Language models pattern-complete. A cut-off CSV looks like a complete one to a model that is not checking, so it continues the pattern with invented rows.

## Edge cases / pitfalls
- Silent truncation producing a valid but short CSV; the row-count check is what catches it.
- The agent helpfully estimating when asked for a number; the instruction must forbid it.
- Partial exports from a previous run being treated as current; timestamp the inputs.
- Unknowns being unacceptable to the requester; then the answer is to fix the export, not to guess.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_Iopitkgrlw7aUuuMQGVDsg