{
  "oags": "0.2",
  "status": "alpha",
  "entry": {
    "id": "IB-arxiv-2607.28287-delegation-highest-rhae",
    "type": "evidence",
    "claim": "On ARC-AGI-3's 25 public games with Claude Opus 4.8 under matched inference budgets, the actor-requested delegation-to-model-builder policy obtained the highest observed mean Relative Human Action Efficiency (RHAE) of 88.49 among four orchestration policies.",
    "evidence_grade": "E3",
    "state": "active",
    "primary_source_ref": [
      {
        "kind": "arxiv",
        "id": "2607.28287",
        "version": "v1",
        "url": "https://arxiv.org/abs/2607.28287",
        "locator": "abstract"
      }
    ],
    "quantities": {
      "mean_rhae": 88.49,
      "public_games": 25,
      "policies_compared": 4
    },
    "declared_loss": null,
    "source_title": "Tycho: Active Abstraction with Programmatic World Models for ARC-AGI-3",
    "authors": [
      "Jens Lehmann",
      "Andrei Aioanei",
      "Sahar Vahdati"
    ],
    "result_origin": "measured",
    "support_quote": "we compare four orchestration policies on all 25 public games using Claude Opus 4.8 under matched inference budgets. Actor-requested delegation to a model builder obtains the highest observed mean Relative Human Action Efficiency (RHAE), 88.49",
    "source_version": "v1",
    "abstract_check": {
      "version": "v1",
      "date": "2026-10-08",
      "quote_verbatim": true
    },
    "domain_v2": "cs-ai/agents"
  },
  "expand": [],
  "incoming": [],
  "links": [],
  "declared_losses": [
    "Public evidence-graph projection: internal analytic/theory edges and private provenance are withheld by design.",
    "Alpha: some primary-source refs await precise citation verification."
  ],
  "generated": "2026-10-08"
}