{
  "schema_version": 2,
  "date": "2026-08-29",
  "scope": [
    "2608.27448-ttpo",
    "2608.27420-weak-guide-rlvr",
    "2608.26735-uc-mopd",
    "2608.26550-spear",
    "2608.27449-swe-prime",
    "2608.27311-harnesslens",
    "2608.26895-covemem",
    "2608.26563-spt"
  ],
  "results": {
    "ttpo": {
      "source": "../post-training/2608.27448-ttpo/metrics/arithmetic-smoke-seed42.json"
    },
    "weak-guide-rlvr": {
      "source": "../post-training/2608.27420-weak-guide-rlvr/metrics/arithmetic-smoke-seed42.json"
    },
    "uc-mopd": {
      "source": "../post-training/2608.26735-uc-mopd/metrics/arithmetic-smoke-seed42.json"
    },
    "spear": {
      "source": "../post-training/2608.26550-spear/metrics/arithmetic-smoke-seed42.json"
    },
    "swe-prime": {
      "source": "../agent-research/2608.27449-swe-prime/metrics/toolroute-l2-seeds42-44.json"
    },
    "harnesslens": {
      "source": "../agent-research/2608.27311-harnesslens/metrics/toolroute-l2-seeds42-44.json"
    },
    "covemem": {
      "source": "../agent-research/2608.26895-covemem/metrics/toolroute-l2-seeds42-44.json"
    },
    "spt": {
      "source": "../agent-research/2608.26563-spt/metrics/toolroute-l2-seeds42-44.json"
    }
  },
  "claim_boundary": "Post-training uses a CPU arithmetic mechanism mini-suite; Agent methods use the no-oracle ToolRoute L2.1 split. Neither substitutes for paper-scale benchmarks.",
  "evaluation_protocol": {
    "tier": "mixed_l1_mechanism_and_l2_1_agent",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "paper results and local public-data evidence remain separate"
  },
  "manifest_ref": "experiments:latest-20260829-seed42",
  "provenance": {
    "artifact_path": "docs/experiments/latest-20260829-seed42.json",
    "dataset_fingerprint": "per-paper artifacts"
  }
}