## TL;DR

The agent classified on weak signals (evening hour, restaurant) and ignored the business evidence. Classify meals on attendees and documented business purpose: a dinner with clients and a stated purpose is business, regardless of the hour.

## Error

```text
agent flagged a legitimate client dinner as personal
```

## Steps

1. Review what signals the agent used for the personal verdict. Expected: you find it weighted time-of-day or venue over attendees.
2. Reweight the classifier: attendees (especially external ones) and a specific business purpose outweigh hour and venue. Expected: business evidence dominates the verdict.
3. Require the agent to cite the evidence for a personal verdict (which signals, what values). Expected: verdicts become explainable.
4. Route low-confidence personal verdicts to manual review instead of auto-flagging. Expected: close calls get human judgment.
5. Feed reviewer overturns back as training corrections. Expected: the classifier learns the business patterns.

## When to use

- Legitimate client dinners get flagged as personal.
- The agent over-weights time or venue signals.

## When not to use

- The meal was genuinely personal (the flag is correct).
- No attendees or purpose are documented (document them first).

## Tool compatibility

- Custom agent expense classifiers; reviewer overturn workflows.

## Variant phrasings

### Business meal misclassified as personal

Same failure, alternate wording.

### False personal-expense flag on client dinner

The false-positive phrasing.

## Why it happens

Personal-vs-business is a judgment call the agent reduces to proxies. Evening restaurants correlate with personal spending in training data, so the agent needs explicit business-evidence signals to override the proxy.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_RhPlK_AiAEu2HmdbQ44fvw