Conflict Detection

Conflict Detection

When the evidence disagrees, the worst thing an AI can do is quietly pick the side that fits its answer. A reader deserves to know the studies conflict — that fact is often the most important thing about the evidence. This skill makes the agent actively compare claims for contradiction and put the conflict on the table.

When to use

After [claim-extraction](../claim-extraction/SKILL.md), before writing any conclusion that leans one way. Feeds [source-reliability](../source-reliability/SKILL.md), which decides how to weigh a conflict once it's found.

The rule

Compare claims that share a topic for directional contradiction. When two claims on the same thing point opposite ways, flag it — do not average it or drop one.

Check each pair of claims that are about the same topic for three kinds of conflict:

1. Directional — "increased" vs "decreased" (or "higher" vs "lower") on the same measure.
2. Effectiveness — "effective" / "benefit" vs "no effect" / "no significant difference" on the same intervention. Check the negated form first, so "not effective" is never misread as "effective".
3. Effect-size direction — a protective estimate (HR/RR/OR below 1) vs a harmful one (above 1) for the same intervention and outcome.

To decide two claims are "about the same topic", compare their topic fingerprint — the meaningful keywords with directional and filler words removed. Two claims only conflict if they genuinely share a subject; a difference on two different outcomes is not a conflict.

Doing it well

• Same topic, opposite direction — that's the test. Different populations or endpoints are not a contradiction; they're context. Don't manufacture conflicts across unrelated claims.
• Handle negation before signals. "No reduction in mortality" carries a negation that flips its meaning; read it before matching directional words.
• Two claims from the same source aren't a conflict. Internal consistency within one chunk is assumed; look for disagreement across sources.
• A real conflict is a finding, not a footnote. Surface it prominently, name both sides with their designs, and let [source-reliability](../source-reliability/SKILL.md) explain which is stronger and why. Never resolve it by averaging.

Output

Conflict on: metformin → all-cause mortality

  S3  cohort (n=12,482)   HR 0.83 [0.76–0.90]   ↓ lower mortality
  S7  RCT   (n=640)       p=0.21, no difference  ● no effect

Type: effectiveness + effect-direction
Note: the RCT (stronger design) shows no effect; the benefit comes from
observational data. Surface both — do not report a single pooled number.

The point is that the reader sees the disagreement. A confident answer that hides a conflict is the failure this skill exists to prevent.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Hash-7777
• Source: Hash-7777/Hash-Medical-Reasearch-Agent-Skills
• License: MIT

Install and usage instructions live in the source repository linked above.