feat: add ranking conflict detection between models

[bledden]

Summary

Adds detection of fundamental disagreements between models based on how they rank each other during Stage 2.

This helps distinguish between:

• Wording differences: Models agree on substance but phrase things differently
• Fundamental conflicts: Models genuinely disagree about response quality

Conflict Types

Mutual Opposition (High Severity)

Both models rank each other poorly while ranking themselves highly. This pattern strongly indicates fundamental disagreement rather than stylistic preferences.

Example: Model A ranks itself #1 and Model B #3, while Model B ranks itself #1 and Model A #3.

Ranking Swap (Medium Severity)

Large position difference in mutual rankings. One model places the other high, but the other places them low.

Example: Model A ranks Model B #1, but Model B ranks Model A #4.

Changes

Backend (backend/council.py):

• New detect_ranking_conflicts() function
• Builds ranking matrix: ranker_rankings[ranker_model][ranked_model] = position
• Detects mutual opposition when both models rank other poorly + self highly
• Detects ranking swaps when position difference ≥ (total_models - 1)
• Returns: model_a, model_b, conflict_type, details, severity

Frontend:

• Stage2.jsx: Added rankingConflicts prop and display component
• Stage2.css: Red-styled card (distinct from yellow minority opinions)
• Severity-colored left border (high=red, medium=orange, low=yellow)
• ChatInterface.jsx: Pass ranking_conflicts to Stage2

Validation

8 test cases in tests/test_ranking_conflicts.py:

Test	Description
test_no_conflict_when_agreement	No false positives when models roughly agree
test_mutual_opposition_detected	Correctly detects mutual opposition pattern
test_ranking_swap_detected	Detects large position disagreements
test_empty_inputs	Handles edge cases gracefully
test_single_model_no_conflict	Single model has no conflicts
test_5_model_conflict_scenario	Realistic 5-model scenario
test_conflict_details_populated	Verifies detail fields are correct
test_severity_ordering	Conflicts sorted by severity

$ python3 tests/test_ranking_conflicts.py
✓ No conflict when agreement - PASSED
✓ Mutual opposition detected - PASSED
✓ Ranking swap detected - PASSED
✓ Empty inputs - PASSED
✓ Single model no conflict - PASSED
  Found 3 conflicts in 5-model scenario
    gpt-4 vs grok: mutual_opposition (high)
    gpt-4 vs llama: mutual_opposition (high)
    grok vs llama: mutual_opposition (high)
✓ 5-model conflict scenario - PASSED
  Details: A ranks B=3, B ranks A=3
  Self ranks: A=1, B=1
✓ Conflict details populated - PASSED
✓ Severity ordering - PASSED

==================================================
All ranking conflict tests passed!
==================================================

Dependencies

This PR builds on:

• PR Add tournament-style pairwise ranking aggregation #128 (tournament ranking)
• PR feat: add minority opinion detection for ranking disagreements #129 (minority opinions)

Test plan

• Run python3 tests/test_ranking_conflicts.py - all 8 tests pass
• Run python3 tests/test_minority_opinions.py - all 8 tests pass
• Start the app and verify conflicts display correctly in Stage 2
• Verify no conflicts shown when models have similar rankings

🤖 Generated with Claude Code