## TL;DR
The agent invented breaking changes and cited release notes that do not exist. Treat every changelog claim as unverified until it links to a real upstream note: fetch the actual releases, quote them, and drop anything that cannot be sourced.

## Error
```text
agent's changelog summary hallucinated breaking changes  -  the release notes it cited didn't exist
```

## Steps
1. Verify the claims: open each cited release note or version in the upstream repository and confirm it exists.
   Expected output: You know which claims are real and which are invented.
2. Fetch the real release notes (repository releases page or changelog file) for the version range of the bump.
   Expected output: You have the actual upstream notes in hand.
3. Rewrite the PR summary to quote only the real notes, with links; remove every unsourced claim.
   Expected output: The summary is short, accurate, and every claim has a source.
4. Change the agent's instructions: changelog sections must be copied from fetched upstream notes, never written from memory, and must include the source link per claim.
   Expected output: Future summaries are verifiable by construction.
5. Re-review the bump itself for real breaking changes the hallucination may have obscured.
   Expected output: You know whether the upgrade is actually safe.

## When to use
- A PR summary cites release notes you cannot find upstream
- The changelog section reads confidently but the versions do not exist
- You are writing summarization rules for an agent

## When not to use
- Real breaking changes in an upgrade; write a migration plan
- Missing changelog sections; that is incompleteness, not hallucination
- Upstream notes that are genuinely sparse; say so instead of inventing

### Variant phrasings
#### upgrade PR cited release notes that do not exist
#### agent invented breaking changes in the changelog summary
#### require sourced changelog claims in dependency PRs

## Why it happens
The agent was asked to summarize changes it never fetched. Language models fill gaps fluently, so it produced plausible-sounding breaking changes with specific-sounding citations. The summary looked authoritative precisely because it was invented.

## Edge cases
- Upstream may keep notes in a changelog file rather than releases; check both before declaring a claim false.
- Pre-release notes sometimes get rewritten; cite the note as of the fetch date.
- A hallucinated summary can still accidentally be right; verify anyway, because the next one will not be.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_Bi8yUdjA5dyBMTIYNXge-A