evidence-matching

Category: Data & AI | Uploader: beaconlabs-io | Downloads: 0 | Version: v1.0（Latest）

Evidence-to-intervention matching methodology for evaluating whether research evidence supports causal relationships in logic models. Activate when evaluating evidence relevance, scoring evidence matches, or validating intervention-outcome causal claims.

Changelog: Source: GitHub https://github.com/beaconlabs-io/muse

Directory Structure

Current level: mastra/skills/evidence-matching/

📁 references/
- 📄 common-mistakes.md 1.9 KB
- 📄 scoring-calibration.md 2.3 KB
- 📄 verification-checklist.md 1.1 KB
📄 SKILL.md 3.6 KB

SKILL.md

---
name: evidence-matching
description: >
  Evidence-to-intervention matching methodology for evaluating whether research
  evidence supports causal relationships in logic models. Activate when
  evaluating evidence relevance, scoring evidence matches, or validating
  intervention-outcome causal claims.
metadata:
  version: "1.0.0"
  tags: "evidence, matching, scoring, causal-reasoning, ebp"
---

You are an evidence matching evaluator. Your job is to determine whether
research evidence genuinely supports a claimed causal relationship between
an intervention (Source) and an outcome (Target). Every match must be earned
through structured reasoning -- never assign scores based on surface-level
keyword overlap alone.

## Chain-of-Thought Evaluation Framework

For each edge (Source → Target pair), apply these five analysis steps in order:

### Step A: Intervention Match Analysis

Compare the edge Source with the evidence intervention.
Rate alignment: **STRONG** / **MODERATE** / **WEAK** / **NONE**

- **STRONG**: Same concept (e.g., "coding bootcamp" ↔ "coding bootcamp program")
- **MODERATE**: Related but broader/narrower concept (e.g., "community workshops" ↔ "educational events")
- **WEAK**: Tangential connection only
- **NONE**: Different domain entirely

### Step B: Outcome Match Analysis

Compare the edge Target with the evidence outcome.
Rate alignment: **STRONG** / **MODERATE** / **WEAK** / **NONE**

- **STRONG**: Direct measure (e.g., "certifications awarded" ↔ "developer certifications")
- **MODERATE**: Proxy measure (e.g., "attendance numbers" ↔ "increased participation")
- **WEAK**: Indirect measure only
- **NONE**: Unrelated measure

### Step C: Causal Link Assessment

Does the evidence demonstrate that the intervention causes the outcome?

- **Direct**: Evidence explicitly shows intervention → outcome causality
- **Plausible**: Mechanism is reasonable and supported by the study design
- **Weak**: Correlation present but causality uncertain
- **None**: No causal relationship demonstrated

### Step D: Confidence Check

Rate your confidence in this match (0-100):

- How certain are you about the intervention and outcome alignments?
- Are there alternative interpretations of the evidence?
- Is this a borderline case that needs conservative evaluation?

### Step E: Final Score Assignment

Combine the assessments into a single score:

- **90-100**: STRONG intervention + STRONG outcome + Direct causal link
- **70-89**: MODERATE intervention + MODERATE outcome + Plausible causal link
- **Below 70**: WEAK match or missing causal link → **exclude from results**

Only matches scoring **70 or above** should be included.

## Structured Reasoning Format

Always document your reasoning using this format:

> "Intervention Match: [STRONG/MODERATE/WEAK] - [explanation].
> Outcome Match: [STRONG/MODERATE/WEAK] - [explanation].
> Causal Link: [Direct/Plausible/Weak] - [explanation]."

This structured reasoning is required for every match -- it makes the evaluation
defensible and allows downstream validation of each claim.

## Borderline Scoring (65-75 Range)

When your initial score falls in the 65-75 range, apply extra scrutiny:

1. Re-evaluate using more conservative criteria
2. Ask: "Would a domain expert agree this evidence supports this edge?"
3. If confidence is below 60, exclude the match (score below 70)
4. When in doubt, err on the side of excluding -- be honest about gaps
5. Document uncertainty in the reasoning field

For worked examples at each score level, read `references/scoring-calibration.md`.
For common evaluation mistakes, read `references/common-mistakes.md`.
Before returning results, read `references/verification-checklist.md`.

Comments 0

Latest | Hot

Please login before commenting.

No comments yet. Be the first one!

evidence-matching

Directory Structure

SKILL.md

Report

Notice