---

name: compare-friction-outputs description: Compare friction pipeline outputs across different runs, thinking levels, or seeds. Use when comparing visa research results, checking variance across thinking levels (minimal/low/medium), analyzing fee discrepancies, or reviewing route selection consistency. allowed-tools: Read, Glob, Grep, Bash

Compare Friction Pipeline Outputs

Compare outputs from the orig_dest_occ_friction pipeline across different runs to analyze variance and consistency.

When to Use

• Comparing results across thinking levels (minimal, low, medium)
• Checking if route selections are consistent
• Investigating field value variance (fees, salaries, processing days)
• Verifying synthesis decisions

Key File Locations

src/orig_dest_occ_friction/
├── artifacts_v8_minimal/
│   └── [triple]/
│       ├── output.json          # FINAL extracted values (use this!)
│       ├── panel.csv            # Yearly panel expansion
│       └── workspace/
│           ├── research.md      # Full research with worker synthesis
│           ├── todos.json       # Field completion status
│           └── year_routes.json # Route selection
├── artifacts_v8_low/
└── artifacts_v8_medium/

Comparison Workflow

1. Find artifact directories

ls -d src/orig_dest_occ_friction/artifacts_*

2. Find common triples

ls src/orig_dest_occ_friction/artifacts_v8_minimal/

3. Compare route selections

For each triple, extract first route from output.json → year_routes.

4. Compare field values

Key fields to compare:

• application_fee_applicant - Total mandatory fees
• salary_min - Minimum salary threshold
• representative_days - Processing time
• pr_years_required - Years to permanent residency
• job_loss_grace_period_days - Grace period after job loss

5. Use the comparison script

uv run python scripts/compare_friction_outputs.py artifacts_v8_minimal artifacts_v8_low artifacts_v8_medium

Understanding research.md Structure

The workspace/research.md file has multiple sections:

1. Worker Reports (top) - 3 parallel workers with potentially different findings
2. Step 1: Decisions - Synthesis rationale explaining which worker was chosen
3. Step 2: Search Queries/Results - Verification searches
4. Step 3: Final Synthesis - THE ACTUAL OUTPUT VALUES (matches output.json)

IMPORTANT: Always compare using output.json or "Step 3: Final Synthesis", NOT the individual worker reports at the top of research.md.

Investigating Variance

When values differ across runs:

1. Read the full research.md for each run
2. Check "Step 1: Decisions" for the rationale
3. Look for which worker was selected and why
4. Check if the variance is due to:
   • Different fee components included
   • Historical data interpretation differences
   • Source availability differences

Output Format

=== ROUTE SELECTIONS ===
Triple                    minimal              low                  medium
australia_china_2211      Subclass 457         Subclass 457         Subclass 457

=== FIELD COMPARISON ===
australia_china_2211:
  ⚠️ application_fee_applicant:  (⚠️ = values differ)
      minimal: 765
      low: 265
      medium: 615
     salary_min:                  (no marker = consistent)
      minimal: 47480
      low: 47480
      medium: 47480