## TL;DR

The agent claimed a matching account that does not exist. Never act on a claimed match without verifying it: read the record back by id from the CRM, and only proceed if the read succeeds and the fields match the claim. Verification is a separate step, not an optimization.

## Error

```text
Verification failed: claimed matching account id 001... returned "record not found" on read-back.
```

## Steps

1. Take the agent's claimed account id and read it directly from the CRM. Expected: either the record or a not-found.
2. If not found, reject the match and log the hallucination with the agent's reasoning. Expected: a recorded miss, not a silent skip.
3. If found, compare key fields (name, domain) against what the agent claimed. Expected: confirmation the right record was meant.
4. Only then proceed with the merge or update. Expected: actions target real records.
5. Add read-back verification as a mandatory gate in the agent loop for every claimed match. Expected: hallucinations caught before damage.

## When to use

- An agent acts on accounts that do not exist.
- Merge or update targets fail with not-found.
- After model changes, hallucinations spike.

## When not to use

- Fuzzy matching accuracy (different tuning task).
- Records deleted between match and action (race, different fix).

## Tool compatibility

- Salesforce and HubSpot record read APIs.
- Agent verification gates.

## Variant phrasings

### phantom account match

Read back by id before acting.

### claimed match not found

Reject and log the hallucination.

## Why it happens

Language models complete patterns plausibly, including inventing ids that look right. Without a grounding read, the plausible fiction flows straight into a destructive call.

## Edge cases

- Read-back must use the id, not a re-search by name; re-searching can hallucinate twice.
- Cache the verification briefly; do not re-verify the same id dozens of times.
- Track hallucination rates per model version; regressions show up here first.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_TuS8CaqRLR9hwduq-YBYaw