bug: sentiment score inaccurate and random on /signals/contacts

[RAWx18]

Short Summary

Sentiment values on the Signals → Contacts page are inaccurate and appear random due to outdated or misconfigured sentiment analysis logic.

Steps to Reproduce

1. Navigate to the Signals tab.
2. Open the Contacts page at /signals/contacts.
3. Inspect the Sentiment column for contacts with comments across platforms.
4. Compare sentiment values against actual comment content.

Expected Behavior

Sentiment should accurately reflect the tone of a contact’s comments using a deterministic and reliable model, producing consistent scores on a 0–100 scale.

Actual Behavior

Sentiment values appear random, inconsistent, and do not correlate with the actual sentiment of the underlying comments.

Environment

• OS: Linux
• Browser: Chrome
• Environment: Local development
• Route: /signals/contacts

Additional Context

Current implementation likely relies on an outdated or legacy Docker-based sentiment model.
Required actions:

• Audit existing sentiment analysis code path and model source.
• Verify whether inference is deterministic and input text is correctly passed.
• Replace with a lightweight, high-accuracy sentiment technique optimized for low memory and token usage.
  Proposed direction:
• Use a small transformer or distilled model or a classical hybrid approach (lexicon + shallow classifier).
• Ensure batch inference and caching to minimize compute overhead.
• Maintain 0–100 output scale with clear normalization logic.
  This fix requires model swap and pipeline refactor, not threshold tuning.