OCG Bench

Code Engineering / Single-File Debug

@code_engineering/single_file_debug

5 transitions4 decisions3 evidence

cost-benefit threshold at ~300 lines / single-call scope

Code Engineering / Large Repo Migration

@code_engineering/large_repo_migration

4 transitions4 decisions3 evidence

file-level decomposition is the only stable path for cross-file state tasks

Code Engineering / Code Review

@code_engineering/code_review

4 transitions4 decisions2 evidence

SecPass 19% across all models; security review always needs human gate

Document Analysis / Long PDF Extraction

@document_analysis/long_pdf_extraction

3 transitions4 decisions3 evidence

multi-needle retrieval is the capability cliff for Sonnet 4.6

Document Analysis / Literature Review

@document_analysis/literature_review

3 transitions3 decisions2 evidence

no model is a reliable source of truth for citation accuracy

Structured Data / JSON Extraction

@structured_data/json_extraction

3 transitions4 decisions2 evidence

schema validation layer is non-negotiable regardless of model

Structured Data / SQL Repair

@structured_data/sql_repair

3 transitions3 decisions2 evidence

engine-specific optimization is the capability threshold between Sonnet and Opus

Language Reasoning / Legal Clause Comparison

@language_reasoning/legal_clause_comparison

4 transitions3 decisions4 evidence

no model is authoritative on binding legal decisions; always human gate

Business Decision / Customer Support Triage

@business_decision/customer_support_triage

3 transitions4 decisions3 evidence

Opus 4.8 cost unjustified for routine triage; Sonnet 4.6 is the correct default

Structured Data / Financial Table Extraction

@structured_data/financial_table_extraction

3 transitions4 decisions4 evidence

numerical verification against source is mandatory regardless of model