Evidence Scorer

The evidence scorer agent turns qualitative support into explicit confidence scores.

What it does

For each claim, it evaluates source quality, methodological rigor, agreement across sources, and reproducibility signals. This produces a transparent ranking of which claims are strong, weak, or uncertain.

The goal is not to replace judgment, but to make judgment auditable and consistent across claims.

Output

A claim-by-claim score table with risk notes and rubric explanation.

Scoring rubric

The evidence scorer typically applies a weighted rubric across:

• Source quality — primary, secondary, or self-reported evidence
• Method quality — controls, sample size, benchmark design, and statistical soundness
• Cross-source consistency — agreement or conflict across independent sources
• Reproducibility signal — availability of code, data, and procedural detail
• Stability — whether the finding is robust across settings and time

This makes low-confidence claims visible early, before they become central to the narrative.

How to interpret scores

Practical interpretation:

• 80-100 — strong support, suitable for high-confidence recommendations
• 50-79 — moderate support, keep caveats explicit
• 0-49 — weak support, treat as exploratory or remove from key conclusions

Thresholds are guidance, not absolute truth; final decisions should consider context and risk tolerance.

When to use it

Use this agent when:

• multiple claims compete for inclusion
• evidence quality varies widely
• stakeholders need transparent confidence framing

It is particularly useful in strategy, product, and research planning documents where weak claims can be costly.

Limitations

A numeric score can hide nuance if read in isolation. Always review accompanying risk notes and contradiction findings before final decisions.

Used by

Most useful in /compare, /deepresearch, and high-stakes decision briefs where prioritization depends on evidence quality.