h-win.json

{
  "description": "H-WIN reference matrix: the HONEST comparison for the headline claim 'a light-tier model inside the system >= a standard-tier model bare' is (light+system) vs (standard+bare) — not bare-vs-system at the same tier. Model aliases (haiku/sonnet) are the tier->alias mapping from roster/model-profiles.yaml; review before a live (non --smoke) run.",
  "arms": [
    {
      "label": "bare-standard",
      "fix_hook": "evals/hooks/keep-bare.sh",
      "env": {
        "EIDOLONS_EVAL_MODEL": "sonnet",
        "EIDOLONS_SANDBOX_MODEL_TIER": "standard"
      },
      "control": true
    },
    {
      "label": "system-light",
      "fix_hook": "evals/hooks/keep-system.sh",
      "env": {
        "EIDOLONS_EVAL_MODEL": "haiku",
        "EIDOLONS_SANDBOX_MODEL_TIER": "light"
      },
      "control": false
    }
  ]
}