docs/implementation/analyst_team_calibration_fixes.md

Analyst Team Calibration Fixes

This update calibrates the existing Analyst Team + Disagreement Resolution layer to reduce false disagreement noise while preserving visibility.

What Was Tightened

• Confidence disagreement now requires spread greater than 0.50.
• DOCUMENT_EVIDENCE disagreement is raised only when document usage is relevant.
• AGGREGATION disagreement is gated by aggregation intent in the question.
• FILTER_LOGIC disagreement is gated by filter intent.
• SCHEMA disagreement is gated by schema/join/link/column ambiguity intent.
• High severity disagreement is now reserved for high-risk final approval scenarios.

New Diagnostics

• DisagreementItem.actionable indicates whether a disagreement should drive behavior.
• Coordinator synthesis now logs:
  • coordinator action vs final action alignment
  • coordinator override flag and reason
  • high-risk approval flag
• Guard diagnostics added:
  • filter_logic_risk
  • schema_join_risk
  • aggregation_risk
  • guard_findings

Lightweight Guards Added

Post-execution, pre-final checks now provide defensive risk signals:

1. Filter guard

• Detects likely mismatched filters when zero rows are returned but nearby values exist.

2. Schema guard

• Flags join/schema risks when join-intent tasks return empty outputs or expected output columns are missing/null-heavy.

3. Aggregation guard

• Flags missing/implausible aggregation behavior when aggregation intent exists.

If required data is unavailable, guard outputs remain unknown instead of failing.

Evaluation Reporting Changes

Verbose/research reports now emphasize actionable signals:

• Tasks With Actionable Disagreement
• High Severity Actionable Disagreements
• Coordinator Override Count
• High Risk Approval Count
• Filter/Schema/Aggregation Guard Risks

Per-task compact columns in verbose/research:

• Agree
• ActDisag
• Risk
• Final
• Coord