# debrief agent summarized interviews but dropped the negative feedback - the sentiment filter ate the critical parts

## TL;DR

A sentiment filter meant to keep summaries 'constructive' silently deleted the critical feedback hiring decisions need. Remove the filter: summaries must preserve concerns verbatim, flagged as concerns. A debrief that hides negatives is a liability, not a kindness.

```text
debrief agent summarized interviews but dropped the negative feedback - sentiment filter
```

## Steps

1. Compare raw interview notes against the summaries for a sample of loops.
   Expected: The negatives are present in notes, absent in summaries.

2. Remove or invert the sentiment filter: concerns must be preserved and labeled, not deleted.
   Expected: Summaries now carry the critical parts.

3. Re-generate summaries for affected loops from the raw notes.
   Expected: The record is corrected.

4. Add a fidelity check: summaries must mention every concern tagged in the notes.
   Expected: Future drops are caught.

5. Review any hiring decisions made on the sanitized summaries.
   Expected: Decisions based on incomplete data are revisited.

## When to use

- summaries are uniformly positive
- known concerns vanish from debriefs
- a sentiment filter touches summaries

## When not to use

- the interviews were genuinely positive (check the raw notes)
- the filter only removes profanity, not concerns (verify what it actually strips)
- summaries are written by humans (different problem)

## Tool compatibility

- any interview-summarization agent
- LLM summarization

## Variant phrasings

- "debrief dropped negative feedback"
- "sentiment filter ate criticism"
- "interview summary too positive"

## Why it happens

Summarization prompts often include 'keep it constructive' or a post-filter that strips negativity, and the model obliges by deleting the exact signal the debrief exists to surface. Hiring managers then decide on bowdlerized evidence. The filter optimized for tone; the business needed truth.

## Edge cases

- preserving concerns does not mean preserving rants; quote the substance neutrally
- the fidelity check needs the concern tags to exist; tag concerns at note-taking time
- revisiting past decisions is sensitive; involve the hiring manager, not just the agent

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_8tiK9HihU4VOcUtDwmHMQA