{
  "schema": "sibyl-public-benchmark-artifacts.v1",
  "generated_at": "2026-07-03T15:15:14.063Z",
  "note": "Whitelisted public reproducibility artifacts. This is not a full repository dump.",
  "artifacts": [
    {
      "title": "Full methodology review",
      "path": "runs/run-2026-06-10-memory-quality-competitive-benchmark-report.md",
      "group": "reports",
      "description": "Claude Code review of memory quality, competitive results, methodology limits, and publication recommendations.",
      "href": "benchmark-artifacts/runs/run-2026-06-10-memory-quality-competitive-benchmark-report.md",
      "exists": true,
      "bytes": 9053
    },
    {
      "title": "Full Claude Code transcript, English",
      "path": "runs/run-2026-06-10-memory-quality-conversation-transcript-en.md",
      "group": "reports",
      "description": "Word-for-word transcript covering benchmark caveats, anti-Sibyl internal tests, and defensible public framing.",
      "href": "benchmark-artifacts/runs/run-2026-06-10-memory-quality-conversation-transcript-en.md",
      "exists": true,
      "bytes": 31299
    },
    {
      "title": "Executive benchmark report",
      "path": "runs/sibyl-memory-benchmark-executive-report-2026-06-07.md",
      "group": "reports",
      "description": "Executive summary of Sibyl, Mem0, Hindsight, Mnemosyne and Phase A runs.",
      "href": "benchmark-artifacts/runs/sibyl-memory-benchmark-executive-report-2026-06-07.md",
      "exists": true,
      "bytes": 8397
    },
    {
      "title": "Token economy run",
      "path": "runs/run-2026-06-04-mcp-run311-token-economy-mega-50q.md",
      "group": "reports",
      "description": "Cost comparison of no memory, full context, and Sibyl retrieval.",
      "href": "benchmark-artifacts/runs/run-2026-06-04-mcp-run311-token-economy-mega-50q.md",
      "exists": true,
      "bytes": 4022
    },
    {
      "title": "Three-arm demo report",
      "path": "runs/run-2026-06-04-sibyl-hackathon-three-arm-demo.md",
      "group": "reports",
      "description": "No memory vs full context vs planner plus Sibyl.",
      "href": "benchmark-artifacts/runs/run-2026-06-04-sibyl-hackathon-three-arm-demo.md",
      "exists": true,
      "bytes": 3191
    },
    {
      "title": "Three-arm demo raw result",
      "path": "runs/run-2026-06-04-sibyl-hackathon-three-arm-demo.raw_result.json",
      "group": "raw results",
      "description": "Raw JSON for the three-arm demo.",
      "href": "benchmark-artifacts/runs/run-2026-06-04-sibyl-hackathon-three-arm-demo.raw_result.json",
      "exists": true,
      "bytes": 5818
    },
    {
      "title": "Sibyl Scale365 retrieval report",
      "path": "runs/run-2026-06-06-sibyl-365d-500c-50q-category-retrieval.md",
      "group": "scale365",
      "description": "Sibyl retrieval report for 500 companies, 365 days, 350 questions.",
      "href": "benchmark-artifacts/runs/run-2026-06-06-sibyl-365d-500c-50q-category-retrieval.md",
      "exists": true,
      "bytes": 1091
    },
    {
      "title": "Sibyl Scale365 retrieval raw result",
      "path": "runs/run-2026-06-06-sibyl-365d-500c-50q-category-retrieval.raw_result.json",
      "group": "scale365",
      "description": "Raw JSON for Sibyl Scale365 retrieval.",
      "href": "benchmark-artifacts/runs/run-2026-06-06-sibyl-365d-500c-50q-category-retrieval.raw_result.json",
      "exists": true,
      "bytes": 678988
    },
    {
      "title": "Sibyl Scale365 Sonnet answer report",
      "path": "runs/run-2026-06-06-sibyl-365d-500c-50q-category-sonnet46.md",
      "group": "scale365",
      "description": "Claude Sonnet answer-only run using Sibyl retrieved context.",
      "href": "benchmark-artifacts/runs/run-2026-06-06-sibyl-365d-500c-50q-category-sonnet46.md",
      "exists": true,
      "bytes": 14700
    },
    {
      "title": "Sibyl Scale365 Sonnet answer raw result",
      "path": "runs/run-2026-06-06-sibyl-365d-500c-50q-category-sonnet46.raw_result.json",
      "group": "scale365",
      "description": "Raw JSON for Sibyl Sonnet answer-only run.",
      "href": "benchmark-artifacts/runs/run-2026-06-06-sibyl-365d-500c-50q-category-sonnet46.raw_result.json",
      "exists": true,
      "bytes": 1042944
    },
    {
      "title": "Mem0 Scale365 retrieval report",
      "path": "runs/run-2026-06-06-mem0-365d-500c-50q-category-option1.md",
      "group": "scale365",
      "description": "Mem0 retrieval report under the tested zero-ingestion-LLM configuration.",
      "href": "benchmark-artifacts/runs/run-2026-06-06-mem0-365d-500c-50q-category-option1.md",
      "exists": true,
      "bytes": 82357
    },
    {
      "title": "Mem0 Scale365 retrieval raw result",
      "path": "runs/run-2026-06-06-mem0-365d-500c-50q-category-option1.raw_result.json",
      "group": "scale365",
      "description": "Raw JSON for Mem0 Scale365 retrieval.",
      "href": "benchmark-artifacts/runs/run-2026-06-06-mem0-365d-500c-50q-category-option1.raw_result.json",
      "exists": true,
      "bytes": 2825931
    },
    {
      "title": "Mem0 Scale365 Sonnet answer report",
      "path": "runs/run-2026-06-06-mem0-365d-500c-sonnet46-answeronly.md",
      "group": "scale365",
      "description": "Claude Sonnet answer-only run using Mem0 retrieved context.",
      "href": "benchmark-artifacts/runs/run-2026-06-06-mem0-365d-500c-sonnet46-answeronly.md",
      "exists": true,
      "bytes": 91964
    },
    {
      "title": "Mem0 Scale365 Sonnet answer raw result",
      "path": "runs/run-2026-06-06-mem0-365d-500c-sonnet46-answeronly.raw_result.json",
      "group": "scale365",
      "description": "Raw JSON for Mem0 Sonnet answer-only run.",
      "href": "benchmark-artifacts/runs/run-2026-06-06-mem0-365d-500c-sonnet46-answeronly.raw_result.json",
      "exists": true,
      "bytes": 3258260
    },
    {
      "title": "Hindsight Scale365 retrieval report",
      "path": "runs/run-2026-06-07-hindsight-365d-500c-50q-category-retrieval.md",
      "group": "scale365",
      "description": "Hindsight retrieval report under the tested no-provider configuration.",
      "href": "benchmark-artifacts/runs/run-2026-06-07-hindsight-365d-500c-50q-category-retrieval.md",
      "exists": true,
      "bytes": 103296
    },
    {
      "title": "Hindsight Scale365 retrieval raw result",
      "path": "runs/run-2026-06-07-hindsight-365d-500c-50q-category-retrieval.raw_result.json",
      "group": "scale365",
      "description": "Raw JSON for Hindsight Scale365 retrieval.",
      "href": "benchmark-artifacts/runs/run-2026-06-07-hindsight-365d-500c-50q-category-retrieval.raw_result.json",
      "exists": true,
      "bytes": 18883857
    },
    {
      "title": "Hindsight Scale365 Sonnet answer report",
      "path": "runs/run-2026-06-07-hindsight-365d-500c-sonnet46-answeronly.md",
      "group": "scale365",
      "description": "Claude Sonnet answer-only run using Hindsight retrieved context.",
      "href": "benchmark-artifacts/runs/run-2026-06-07-hindsight-365d-500c-sonnet46-answeronly.md",
      "exists": true,
      "bytes": 111114
    },
    {
      "title": "Hindsight Scale365 Sonnet answer raw result",
      "path": "runs/run-2026-06-07-hindsight-365d-500c-sonnet46-answeronly.raw_result.json",
      "group": "scale365",
      "description": "Raw JSON for Hindsight Sonnet answer-only run.",
      "href": "benchmark-artifacts/runs/run-2026-06-07-hindsight-365d-500c-sonnet46-answeronly.raw_result.json",
      "exists": true,
      "bytes": 19327228
    },
    {
      "title": "Mnemosyne Scale365 retrieval report",
      "path": "runs/run-2026-06-07-mnemosyne-365d-500c-50q-category-retrieval.md",
      "group": "scale365",
      "description": "Mnemosyne retrieval report under the tested FTS-only configuration.",
      "href": "benchmark-artifacts/runs/run-2026-06-07-mnemosyne-365d-500c-50q-category-retrieval.md",
      "exists": true,
      "bytes": 74616
    },
    {
      "title": "Mnemosyne Scale365 retrieval raw result",
      "path": "runs/run-2026-06-07-mnemosyne-365d-500c-50q-category-retrieval.raw_result.json",
      "group": "scale365",
      "description": "Raw JSON for Mnemosyne Scale365 retrieval.",
      "href": "benchmark-artifacts/runs/run-2026-06-07-mnemosyne-365d-500c-50q-category-retrieval.raw_result.json",
      "exists": true,
      "bytes": 2971071
    },
    {
      "title": "Mnemosyne Scale365 Sonnet answer report",
      "path": "runs/run-2026-06-07-mnemosyne-365d-500c-sonnet46-answeronly.md",
      "group": "scale365",
      "description": "Claude Sonnet answer-only run using Mnemosyne retrieved context.",
      "href": "benchmark-artifacts/runs/run-2026-06-07-mnemosyne-365d-500c-sonnet46-answeronly.md",
      "exists": true,
      "bytes": 78591
    },
    {
      "title": "Mnemosyne Scale365 Sonnet answer raw result",
      "path": "runs/run-2026-06-07-mnemosyne-365d-500c-sonnet46-answeronly.raw_result.json",
      "group": "scale365",
      "description": "Raw JSON for Mnemosyne Sonnet answer-only run.",
      "href": "benchmark-artifacts/runs/run-2026-06-07-mnemosyne-365d-500c-sonnet46-answeronly.raw_result.json",
      "exists": true,
      "bytes": 3373201
    },
    {
      "title": "Honcho vs Sibyl README",
      "path": "runs/honcho-vs-sibyl-42k-baseline/00-README.md",
      "group": "honcho baseline",
      "description": "Index for the 42k Honcho vs Sibyl baseline bundle.",
      "href": "benchmark-artifacts/runs/honcho-vs-sibyl-42k-baseline/00-README.md",
      "exists": true,
      "bytes": 1893
    },
    {
      "title": "Sibyl 42k Sonnet report",
      "path": "runs/honcho-vs-sibyl-42k-baseline/01-sibyl-42k-sonnet-report.md",
      "group": "honcho baseline",
      "description": "Sibyl side of the 42k Honcho baseline.",
      "href": "benchmark-artifacts/runs/honcho-vs-sibyl-42k-baseline/01-sibyl-42k-sonnet-report.md",
      "exists": true,
      "bytes": 17696
    },
    {
      "title": "Sibyl 42k Sonnet raw result",
      "path": "runs/honcho-vs-sibyl-42k-baseline/02-sibyl-42k-sonnet-raw_result.json",
      "group": "honcho baseline",
      "description": "Raw JSON for the Sibyl 42k run.",
      "href": "benchmark-artifacts/runs/honcho-vs-sibyl-42k-baseline/02-sibyl-42k-sonnet-raw_result.json",
      "exists": true,
      "bytes": 890984
    },
    {
      "title": "Honcho 42k Sonnet report",
      "path": "runs/honcho-vs-sibyl-42k-baseline/03-honcho-42k-sonnet-baseline-report.md",
      "group": "honcho baseline",
      "description": "Honcho side of the 42k baseline.",
      "href": "benchmark-artifacts/runs/honcho-vs-sibyl-42k-baseline/03-honcho-42k-sonnet-baseline-report.md",
      "exists": true,
      "bytes": 32801
    },
    {
      "title": "Honcho 42k Sonnet raw result",
      "path": "runs/honcho-vs-sibyl-42k-baseline/04-honcho-42k-sonnet-baseline-raw_result.json",
      "group": "honcho baseline",
      "description": "Raw JSON for the Honcho 42k run.",
      "href": "benchmark-artifacts/runs/honcho-vs-sibyl-42k-baseline/04-honcho-42k-sonnet-baseline-raw_result.json",
      "exists": true,
      "bytes": 2153634
    },
    {
      "title": "Honcho baseline runner",
      "path": "runs/honcho-vs-sibyl-42k-baseline/05-honcho-baseline-runner.py",
      "group": "scripts",
      "description": "Script used to generate the Honcho 42k baseline.",
      "href": "benchmark-artifacts/runs/honcho-vs-sibyl-42k-baseline/05-honcho-baseline-runner.py",
      "exists": true,
      "bytes": 30813
    },
    {
      "title": "Sibyl Scale365 runner",
      "path": "scripts/run_sibyl_365d_500c_category_benchmark.py",
      "group": "scripts",
      "description": "Sibyl Scale365 benchmark runner.",
      "href": "benchmark-artifacts/scripts/run_sibyl_365d_500c_category_benchmark.py",
      "exists": true,
      "bytes": 33289
    },
    {
      "title": "Mem0 Scale365 runner",
      "path": "scripts/run_mem0_365d_500c_category_baseline.py",
      "group": "scripts",
      "description": "Mem0 Scale365 benchmark runner.",
      "href": "benchmark-artifacts/scripts/run_mem0_365d_500c_category_baseline.py",
      "exists": true,
      "bytes": 25512
    },
    {
      "title": "Hindsight Scale365 runner",
      "path": "scripts/run_hindsight_365d_500c_category_baseline.py",
      "group": "scripts",
      "description": "Hindsight Scale365 benchmark runner.",
      "href": "benchmark-artifacts/scripts/run_hindsight_365d_500c_category_baseline.py",
      "exists": true,
      "bytes": 22634
    },
    {
      "title": "Mnemosyne Scale365 runner",
      "path": "scripts/run_mnemosyne_365d_500c_category_baseline.py",
      "group": "scripts",
      "description": "Mnemosyne Scale365 benchmark runner.",
      "href": "benchmark-artifacts/scripts/run_mnemosyne_365d_500c_category_baseline.py",
      "exists": true,
      "bytes": 21229
    },
    {
      "title": "Answer-only runner",
      "path": "scripts/run_raw_memory_answer_llm.py",
      "group": "scripts",
      "description": "Shared Claude Sonnet answer-only runner for retrieved contexts.",
      "href": "benchmark-artifacts/scripts/run_raw_memory_answer_llm.py",
      "exists": true,
      "bytes": 13352
    },
    {
      "title": "Three-arm demo runner",
      "path": "scripts/run_sibyl_hackathon_three_arm_demo.py",
      "group": "scripts",
      "description": "Runner for the no-memory, full-context, planner plus Sibyl demo.",
      "href": "benchmark-artifacts/scripts/run_sibyl_hackathon_three_arm_demo.py",
      "exists": true,
      "bytes": 14476
    }
  ]
}