{
  "schema": "bia.vibe-benchmark-protocol/v1",
  "protocol_id": "vibe-coding-full-stack-v1",
  "title": "Autonomous full-stack delivery benchmark",
  "language": "pt-BR",
  "providers": [
    "bia",
    "lovable",
    "v0",
    "bolt",
    "replit"
  ],
  "minimum_repetitions_per_prompt_provider": 5,
  "maximum_trial_seconds": 1800,
  "result_validity_days": 90,
  "publication": {
    "publish_every_planned_cell": true,
    "publish_failures_and_timeouts": true,
    "allow_replacement_of_completed_run": false,
    "require_raw_trial_records": true,
    "require_exact_product_versions": true,
    "require_limitations": true
  },
  "operator_actions_not_counted_as_intervention": [
    "create_blank_project",
    "submit_canonical_prompt_once",
    "accept_a_platform_native_permission_required_to_continue",
    "click_the_single_standard_publish_action"
  ],
  "interventions_counted": [
    "answer_clarifying_question",
    "send_followup_prompt",
    "edit_generated_code_or_data",
    "retry_or_regenerate_manually",
    "change_model_or_generation_settings",
    "debug_or_reconfigure_runtime",
    "repair_or_replace_a_failed_deploy"
  ],
  "trial_terminal_statuses": [
    "completed",
    "failed",
    "timeout"
  ],
  "failure_codes": [
    "build_failed",
    "boot_failed",
    "functional_failed",
    "human_intervention",
    "provider_error",
    "provider_limit",
    "timeout",
    "cost_unavailable"
  ],
  "cost": {
    "currency": "USD",
    "unit": "micros",
    "never_treat_unavailable_as_zero": true,
    "accepted_sources": [
      "provider_usage",
      "provider_invoice",
      "metered_plan_allocation"
    ]
  },
  "fairness": {
    "new_blank_project_per_trial": true,
    "same_prompt_bytes_for_every_provider": true,
    "no_private_preview_features": true,
    "default_model_when_provider_does_not_expose_choice": true,
    "randomize_execution_order": true,
    "blind_functional_evaluation": true,
    "one_run_cannot_omit_or_replace_a_planned_trial": true
  }
}