context-controls/initial/experiment.json

{
  "schema": "evalarc.skill-context-experiment.v1",
  "model": {
    "model": "Qwen/Qwen3-8B",
    "revision": "b968826d9c46dd6066d109eabc6255188de91218",
    "device": "NVIDIA L40S",
    "dtype": "bfloat16",
    "torch": "2.11.0+cu130",
    "transformers": "5.5.4",
    "thinking_enabled": false
  },
  "conditions": [
    "relevant",
    "neutral"
  ],
  "route": "mcp",
  "model_seeds": [
    17,
    41,
    97
  ],
  "evaluation_seeds": [
    41,
    97
  ],
  "task_context": "inline",
  "max_steps": 12,
  "max_new_tokens_per_step": 4096,
  "wall_seconds": 600,
  "temperature": 0.2,
  "model_visible_open_skill_tokens": 476,
  "match_sha256": "cb9c2b97a1b056afc8479449b8c0bc8828e2d65e58a9115518ede94012b82ff6",
  "harness_files": {
    "record_skill_impact.py": "804188699a8267295e9d0a4f1498e9892b945766aa919f3e08feff39380f77a1",
    "bridge.mjs": "adaa2b517a367c063eeda25a32a47f232642b0a254718692581cc2a816e81960",
    "record_context_controls.py": "dc5908b432f0ba02000c13ec2184635f8cb86d0fb1a8bafd9123d16e77bdfcad",
    "runner.py": "53346d1eaec035586e9a489b6751654300453c3669c313dca4402a36f4620f5f",
    "agent_sandbox.py": "f949e3b3536ea11a6457e062867c003a1d180609638166c11fe92914e0d72422",
    "robot_task.py": "f8f2fb3e4a163bb35315ae5a97376231eba041d0af0193ecaa776530912730cc",
    "local_model_server.py": "94125874cd72aef26c3487761c2841323ec87812bf10ccb268615565be66af59"
  },
  "results_policy": "All six scheduled trials retained, including errors; no retries.",
  "scope": "One public development task; a new paired context-content control. Same model, catalog metadata, tools, task and budgets. First skill-load payloads have identical token lengths; all later token usage is measured. No independent-author or held-out evaluation, and no pooled efficacy claim."
}