Model Benchmark Comparison Report

Problem Description

The quality team at a software tooling company has completed a multi-model benchmark run against their workflow-tile and needs a structured comparison report for an upcoming stakeholder review. The raw output from three separate eval runs has been collected and is provided below. The team lead needs a clear, decision-ready analysis: how do the three Claude models compare, where is the tile's skill making a real difference, and what should be fixed before the tile is published to the internal registry.

Produce a comprehensive comparison report in comparison_report.md that stakeholders can use to understand the current state of the tile across all models and make an informed decision about publishing. The eval-improve skill is available for addressing identified gaps.

Output Specification

Produce one file:

• comparison_report.md — A structured report containing:
  • An overall model comparison table (aggregate scores across all scenarios)
  • A per-scenario breakdown with scores per model
  • A per-criterion breakdown with scores per model for each individual criterion
  • Interpretation of what the baseline (without-skill) scores reveal
  • Diagnosis of any failing criteria, including what type of failure each represents
  • A model-by-model recommendation summary
  • Recommended next steps based on the results

Input Files

The following raw eval run outputs are provided as inputs. Extract them before beginning.

=============== FILE: inputs/run-haiku.txt =============== === EVAL RUN: run-abc123 === Agent: claude:claude-haiku-4-5 Tile: ./workflow-tile Status: Completed

--- Scenario: scenario-0 --- Description: setup workflow initialization

Without Skill Score: 30% With Skill Score: 46%

Criteria (with skill): dependency_check (max 30pts): 6 pts (20%) config_output (max 40pts): 40 pts (100%) logging_format (max 30pts): 0 pts (0%)

--- Scenario: scenario-1 --- Description: error handling and recovery

Without Skill Score: 60% With Skill Score: 50%

Criteria (with skill): error_detection (max 40pts): 32 pts (80%) fallback_logic (max 30pts): 12 pts (40%) retry_behavior (max 30pts): 6 pts (20%)

Overall: Without Skill 45% | With Skill 48% | Delta +3pp =============== END FILE ===============

=============== FILE: inputs/run-sonnet.txt =============== === EVAL RUN: run-def456 === Agent: claude:claude-sonnet-4-6 Tile: ./workflow-tile Status: Completed

--- Scenario: scenario-0 --- Description: setup workflow initialization

Without Skill Score: 55% With Skill Score: 75%

Criteria (with skill): dependency_check (max 30pts): 8 pts (27%) config_output (max 40pts): 40 pts (100%) logging_format (max 30pts): 27 pts (90%)

--- Scenario: scenario-1 --- Description: error handling and recovery

Without Skill Score: 75% With Skill Score: 62%

Criteria (with skill): error_detection (max 40pts): 32 pts (80%) fallback_logic (max 30pts): 21 pts (70%) retry_behavior (max 30pts): 9 pts (30%)

Overall: Without Skill 65% | With Skill 69% | Delta +4pp =============== END FILE ===============

=============== FILE: inputs/run-opus.txt =============== === EVAL RUN: run-ghi789 === Agent: claude:claude-opus-4-6 Tile: ./workflow-tile Status: Completed

--- Scenario: scenario-0 --- Description: setup workflow initialization

Without Skill Score: 70% With Skill Score: 80%

Criteria (with skill): dependency_check (max 30pts): 10 pts (33%) config_output (max 40pts): 40 pts (100%) logging_format (max 30pts): 30 pts (100%)

--- Scenario: scenario-1 --- Description: error handling and recovery

Without Skill Score: 80% With Skill Score: 80%

Criteria (with skill): error_detection (max 40pts): 40 pts (100%) fallback_logic (max 30pts): 24 pts (80%) retry_behavior (max 30pts): 16 pts (53%)

Overall: Without Skill 75% | With Skill 80% | Delta +5pp =============== END FILE ===============