[divijagv]

What this adds

A synthetic example, in the dataset's trace format, of how a chain of AI agents can
be misled by forged text. An intake agent passes an outside party's claim to a triage
agent as if it were a verified fact, so a high-risk fraud case is closed without a
human ever seeing it.

It is data, not code: one new JSON entry plus the index and checksum updates that go
with it, so merging it does not change any software. Researchers and red teams can use
it to test whether multi-agent systems keep track of where each piece of information
came from. The entry lists each failure point span by span, with suggested mitigations.