Data Analysis

Guidelines for analyzing structured data and producing insights.

Input Handling

• Use ReadArtifact to get dependency outputs (e.g. benchmark results, survey data, configs).
• Use ReadFile for sandbox/* or plan-level files.
• Parse JSON or tabular data. Identify structure (keys, arrays, nested objects).
• Validate structure first: Check for expected keys, types. Log structure if unclear.

Analysis Types

Type	Input	Output
Aggregate	List of items	Metrics (count, sum, avg, min, max), distribution
Compare	Multiple datasets	Side-by-side comparison, deltas
Trend	Time-series or ordered data	Patterns, trends, anomalies
Correlation	Multi-variate data	Relationships, dependencies
Summary	Large dataset	Key statistics, representative samples
Distribution	Categorical/numerical	Frequency table, percentiles
Ranking	Scored items	Ordered list, top-N, thresholds

Output Format

Markdown Report

• Methodology: What was analyzed, data sources.
• Results: Key metrics, tables, findings.
• Interpretation: What the numbers mean.
• Conclusions: Takeaways, recommendations.

JSON Output

{
  "metrics": {"count": N, "avg": X, "min": Y, "max": Z},
  "findings": ["finding1", "finding2"],
  "recommendations": ["rec1", "rec2"]
}

• Match the output spec exactly when format is specified.
• Use consistent key names as defined in task output spec.

Tables for Results

Use Markdown tables for numerical results:

| Metric | Value A | Value B | Delta |
|--------|---------|---------|-------|
| Latency | 10ms | 15ms | +50% |

• Include units in header or first row when relevant.
• Use alignment row (|---|) for readability.

Sandbox Usage

• Save intermediate parsed data to sandbox/parsed.json
• Save computed metrics to sandbox/metrics.json
• Final output via Finish (do not output inline for JSON/Markdown spec)

Error Handling

• Malformed input: Document the issue, suggest correction. Proceed with partial data if possible.
• Missing data: Note gaps explicitly in report; state assumptions.
• Empty input: Return structured empty result with explanation, do not fail silently.
• Type mismatch: Coerce or skip; document in methodology.

Edge Cases

• Single-item list: Still compute metrics; avoid division-by-zero.
• Nested arrays: Flatten or aggregate per level as appropriate.
• Null/missing values: Exclude from numeric aggregates; count separately if relevant.