Contradiction

Run adversarial contradiction analysis against a claim or draft.

The contradiction workflow pressure-tests conclusions by looking for direct conflicts, boundary failures, and hidden assumptions.

How it works

The contradiction workflow runs as an adversarial pass:

  1. Claim extraction — identifies key assertions from your draft or claim statement.
  2. Counter-evidence search — looks for conflicting findings, alternate explanations, and edge cases.
  3. Severity tagging — classifies issues as fatal, major, or minor.
  4. Resolution mapping — proposes concrete fixes (rewrite, verify, or remove).

This helps prevent overconfident conclusions from shipping without stress testing.

Real examples

Example 1 (draft stress test):

/contradict outputs/my-brief.md
bohr contradict outputs/my-brief.md

Example 2 (single claim challenge):

/contradict "RAG always beats fine-tuning for support QA"
bohr contradict "RAG always beats fine-tuning for support QA"

Expected output

Primary output file:

  • outputs/<slug>-contradictions.md

Typical structure:

  • challenged claims list
  • contradiction entries with evidence and severity
  • fragile assumptions summary
  • remediation checklist

Example snippet:

## Contradiction Log
- [C2] Claim overgeneralizes benchmark result to production.
  - Severity: Major
  - Evidence: Source [4] reports gains only under controlled retrieval quality.
  - Resolution: Narrow wording and add production caveat.

Customization

You can tune this workflow by asking for:

  • strictness — “treat unsupported quantitative claims as fatal”
  • scope — “only test claims in executive summary”
  • evidence policy — “prefer peer-reviewed sources over blog evidence”
  • output style — “give rewrite-ready fix language per issue”

When to use

Run this before sharing external reports, high-stakes recommendations, or publication drafts.

For broad discovery work, run /deepresearch first and use /contradict near the end.