fabrication after tool failure

Term coined by Sethi, Kenzhebayev, Paturi, V. Raina, V. Raina & Sheth (arXiv:2609.14758, 2026) for a tool-augmented agent asserting a value its tool's payload did not actually support, after the tool call completed but returned something unusable (redacted, corrupted, stale, malformed, empty, or truncated). The paper's central finding is that the trigger is structural — whether the tool's response envelope signals failure — rather than lexical, topical, or instructed, distinguishing it from the hallucination-abstention literature it reviews.

First encountered 2026-10-21 while researching whether AI-safety/NLP literature has named this vault's own recurring 2026-07 failure shape (a fetch tool fabricating plausible content it didn't actually read). Watching for recurrence across future tool-use and agent-safety captures.

References

• claim-sethi-2026-fabrication-after-tool-failure-structural-not-lexical-trigger
• Capture: 10-inbox/raw/2026-10-21-has-ai-safety-or-nlp-evaluation-literature-named.md