{
  "version": "1.0.0",
  "description": "Execution configuration for the data-collection runner (run/run-experiment.js). A 'cell' is one combination of model x task x language x temperature; 'reps' completions are collected per cell. Change 'reps' only between runs, never inside one (the resume logic keys on rep index).",
  "request": {
    "description": "Per-request settings. max_tokens is deliberately tiny — answers are single words; anything longer is coded invalid anyway. Set to 16 (from 12) because OpenAI chat models reject max_output_tokens<16; 16 is their documented minimum and does not change any observed distribution (valid answers are 1-6 tokens). This is a deviation from the OSF preregistration (which stated 12), documented in paper1. top_p left at default 1.0.",
    "max_tokens": 16,
    "timeout_ms": 90000,
    "concurrency": 16,
    "max_retries": 5,
    "backoff_base_ms": 2000,
    "backoff_max_ms": 60000
  },
  "temperatures": {
    "description": "Sampling plan. t=1.0 estimates the response DISTRIBUTION (the fingerprint); t=0.0 gives a deterministic fingerprint — 3 reps to verify determinism, not to sample.",
    "pilot": [
      { "t": 1.0, "reps": 20 },
      { "t": 0.0, "reps": 3 }
    ],
    "main": [
      { "t": 1.0, "reps": 30 },
      { "t": 0.0, "reps": 3 }
    ]
  },
  "pilot": {
    "description": "Pilot scope (v3). Model slugs MUST be validated against data/models-catalog.json (npm run assess-models warns about unknown slugs). Spans 5 distinct families at low cost, incl. three co-family large+small pairs (nemotron, qwen, deepseek) so intra- vs inter-family separation is estimable from the pilot. HISTORY: pilot-01 (':free' twins) hit OpenRouter's account-level free-models-per-day cap; pilot-02 (paid twins) surfaced that paid endpoints enable hidden reasoning by default, burning the 12-token budget (empty completions) — fixed by reasoning:{enabled:false} in lib.js, and mandatory-reasoning models (gpt-oss pair) were excluded study-wide (PI decision, see exclusion rules); replaced by the qwen *-2507 instruct pair. See data/runs/pilot-0{1,2}/manifest.json notes.",
    "models": [
      "nvidia/nemotron-3-ultra-550b-a55b",
      "nvidia/nemotron-3-nano-30b-a3b",
      "qwen/qwen3-235b-a22b-2507",
      "qwen/qwen3-30b-a3b-instruct-2507",
      "deepseek/deepseek-v4-pro",
      "deepseek/deepseek-v4-flash",
      "meta-llama/llama-3.3-70b-instruct",
      "google/gemma-4-31b-it"
    ],
    "tasks": "all",
    "languages": ["en", "ru", "zh", "ar"]
  },
  "main": {
    "description": "Main run scope. Models come from config/models.selected.json (produced by npm run assess-models, then manually reviewed — set 'included': false to drop a model). Frontier-priced models (see 'expensive_input_threshold_usd_per_mtok') get reduced reps to control cost.",
    "models_file": "config/models.selected.json",
    "tasks": "all",
    "languages": ["en", "ru", "zh", "ar"],
    "expensive_input_threshold_usd_per_mtok": 5.0,
    "expensive_reps_factor": 0.5
  },
  "provenance": {
    "description": "Recorded into each run's manifest.json for reproducibility.",
    "record_git_commit": true,
    "record_prompts_hash": true,
    "record_config_hash": true
  }
}