Design robust contradiction-detection signal for research pipeline

[TimSimpsonJr]

Background

The multi-hop research pipeline rework (see docs/plans/2026-05-26-superclaude-methodology-rework.md and the design doc landing alongside it) introduces a contradiction_rate quality signal that triggers replanning when sources disagree above a threshold.

For v1, contradiction detection is folded into the classify-agent's prompt — it compares key_claims across summaries and flags semantic conflicts. That's cheap but soft: depends on LLM judgment, sensitive to prompt drift, hard to test offline.

Goal

Design a more robust mechanism to replace the v1 in-classify approach. Output is a design memo in docs/plans/, not implementation.

Candidates to weigh

• Dedicated contradiction-scanner Haiku agent. Runs after summarize. Does pairwise comparison of key_claims across all summaries with structured JSON output. Pros: isolated, testable, swappable. Cons: extra agent call per run.
• Embedding-based semantic conflict detection. Compute embeddings of each claim, find pairs with high topical similarity but opposing stance markers. Pros: cheap, deterministic, offline-friendly. Cons: requires an embedding model in the tier stack; stance detection is its own hard problem.
• Citation-graph analysis. Track which sources back which claims; flag when authoritative sources (T1/T2) cite contradictory positions. Pros: ties contradiction to source authority directly. Cons: requires structured claim-to-source provenance we don't track today.
• Temporal authority weighting. When sources from different eras contradict, weight newer sources more for civic/legislative topics. Pros: matches the domain. Cons: not a detection mechanism on its own — a resolution mechanism. Pair with one of the above.

Acceptance criteria

• Design memo with recommended approach, why, and rough implementation cost
• Trade-offs noted vs. the v1 in-classify approach
• Specific contract: how the signal is exposed to the quality-gate (numeric rate? per-claim flags? both?)
• Test plan: how to verify contradiction detection offline

Out of scope

• Implementation
• Choice of embedding model (defer to the recommendation)

Related

• Design doc: docs/plans/2026-05-26-superclaude-methodology-rework.md (and forthcoming design doc with full rework spec)