{
  "benchmark": "Maestro Brief Django customer-request portal",
  "scoring_date": "2026-08-10",
  "acceptance_checks_per_run": 23,
  "runs": [
    {"provider":"Anthropic","model":"Claude Fable 5","run":1,"accepted":23,"total":23,"manual_quality":96.3,"reported_elapsed_minutes":12.0,"reported_agent_tokens":null,"token_note":"Clean per-run receipt unavailable"},
    {"provider":"Anthropic","model":"Claude Fable 5","run":2,"accepted":23,"total":23,"manual_quality":96.1,"reported_elapsed_minutes":13.5,"reported_agent_tokens":128497},
    {"provider":"Anthropic","model":"Claude Opus 5","run":1,"accepted":22,"total":23,"manual_quality":96.6,"reported_elapsed_minutes":21.7,"reported_agent_tokens":158284,"result_note":"Raw miss was a literal TODO inside the repository's own hygiene test; portal behavior passed"},
    {"provider":"Anthropic","model":"Claude Opus 5","run":2,"accepted":23,"total":23,"manual_quality":96.5,"reported_elapsed_minutes":18.3,"reported_agent_tokens":146001},
    {"provider":"Anthropic","model":"Claude Sonnet 5","run":1,"accepted":21,"total":23,"manual_quality":86.5,"reported_elapsed_minutes":13.9,"reported_agent_tokens":131523,"result_note":"Assignment and status activity entries were lost"},
    {"provider":"Anthropic","model":"Claude Sonnet 5","run":2,"accepted":22,"total":23,"manual_quality":89.3,"reported_elapsed_minutes":14.9,"reported_agent_tokens":132496,"result_note":"Administrator omitted from seeded request ownership"},
    {"provider":"OpenAI","model":"Codex Sol","run":1,"accepted":23,"total":23,"manual_quality":90.0,"reported_elapsed_seconds":244,"observed_cache_api_equivalent_usd":0.66},
    {"provider":"OpenAI","model":"Codex Sol","run":2,"accepted":23,"total":23,"manual_quality":90.8,"reported_elapsed_seconds":288,"observed_cache_api_equivalent_usd":0.55},
    {"provider":"OpenAI","model":"Codex Terra","run":1,"accepted":23,"total":23,"manual_quality":89.2,"reported_elapsed_seconds":222,"observed_cache_api_equivalent_usd":0.20},
    {"provider":"OpenAI","model":"Codex Terra","run":2,"accepted":23,"total":23,"manual_quality":89.8,"reported_elapsed_seconds":298,"observed_cache_api_equivalent_usd":0.33},
    {"provider":"OpenAI","model":"Codex Luna","run":1,"accepted":23,"total":23,"manual_quality":88.5,"reported_elapsed_seconds":331,"observed_cache_api_equivalent_usd":0.04},
    {"provider":"OpenAI","model":"Codex Luna","run":2,"accepted":22,"total":23,"manual_quality":83.8,"reported_elapsed_seconds":395,"observed_cache_api_equivalent_usd":0.06,"result_note":"Administrator omitted from seeded ownership; extra debug user and request remained"}
  ],
  "caveats": [
    "Claude builds ran concurrently; Codex builds ran serially",
    "Claude aggregate agent tokens are not comparable to Codex fresh, cached, and output accounting",
    "No API-equivalent Claude cost was calculated",
    "Manual quality scoring is separate from the objective acceptance result",
    "One greenfield Django task cannot establish a universal model ranking"
  ]
}