{
  "schema_version": 1,
  "checked_date_utc": "2026-09-21",
  "measurement_started_at_utc": "2026-09-21T04:43:56.339953+00:00",
  "recorded_workload": {
    "capture_date_utc": "2026-09-21",
    "kunz_simultaneous_researchers_observed": 49,
    "biology_concurrent_campaigns_observed": 30,
    "concurrency_source": "Separate native fleet captures, not transcript counts alone.",
    "recordings": [
      {
        "domain": "kunz",
        "url": "/media/kunz-live.mp4",
        "duration_seconds": 45
      },
      {
        "domain": "biology",
        "url": "/media/biology-live.mp4",
        "duration_seconds": 45
      }
    ],
    "archived_research_sessions": 79,
    "kunz_research_sessions": 49,
    "biology_research_sessions": 30,
    "separate_runtime_probe_sessions": 1,
    "total_archived_transcripts": 80,
    "output_markdown_documents": 188,
    "draft_documents": 79,
    "follow_up_self_audit_documents": 79,
    "launch_audit_documents": 30,
    "controlling_source_documents_outside_output_count": 1,
    "review_scope": "Follow-up self-audits in each original agent session.",
    "count_method": "Match archived bindings to distinct transcript identities. Count Markdown only within campaign output directories. Confirm one draft and one self-audit per research binding.",
    "archive_sha256": "8b34ba2c10f6f002baeef05be10cde2602da65350ab9f546cafe1335f9a84140",
    "scope": "Native workload observations from the two 45-second recordings."
  },
  "controlled_system_checks": {
    "runtime": "forge-research 0.4.8",
    "binary_kind": "existing Windows release executable",
    "binary_sha256": "7d3acaf1dfbd5241101d84177bc4a66d6d2bf8539e81a6519fadacb8e9017248",
    "environment": {
      "operating_system": "Windows 11",
      "build": "26200",
      "python": "3.14.0"
    },
    "fixture": {
      "targets": 1,
      "claims": 2,
      "dependency_edges": 1,
      "evidence_artifacts": 1,
      "strictly_discharged_obligations": 1,
      "synthetic_final_review_identities": 4
    },
    "fresh_process_context_reopens": {
      "trials": 20,
      "identical_context_results": 20,
      "comparison": "Complete JSON response, canonicalized by sorting object keys.",
      "transcripts_included": false,
      "process_exit_condition": "normal exits",
      "cache_condition": "warm operating-system cache"
    },
    "completion_control": {
      "result": "passed",
      "observed_state": "COMPLETE",
      "method": "Complete a closed-workspace clone after four synthetic candidate-bound review passes."
    },
    "evidence_withdrawal": {
      "result": "passed",
      "completion_rejected": true,
      "rejection_reason": "candidate artifact has no current verified authority",
      "claims_remained_active_before_explicit_retraction": true
    },
    "premise_retraction": {
      "controlled_cases": 1,
      "postconditions_checked": 4,
      "postconditions_passed": 4,
      "premise_validity": "RETRACTED",
      "dependent_validity": "INVALIDATED_BY_DEPENDENCY",
      "obligation_status": "OPEN",
      "obligation_discharge_basis": "NONE",
      "goal_state": "ACTIVE"
    },
    "model_calls": 0,
    "daemon_launches": 0,
    "cluster_jobs": 0,
    "script_sha256": "6956c266ec7c6494fd6f0df5c1638204a6684012b6781dc356a2d4c144df0b37",
    "complete_local_result_sha256": "db2d79e55296505e11391723bb03a91f10bfcb21ff66fb64cab69cdfc9cc34b0",
    "conditions": [
      "One synthetic two-claim fixture.",
      "Normal process exits and a warm operating-system cache.",
      "Four retraction postconditions in one controlled case.",
      "Synthetic review identities exercise completion policy.",
      "The binary hash identifies the tested Windows executable."
    ]
  },
  "research_status": {
    "as_of": "2026-09-21",
    "works": [
      {
        "topic": "Cactus regularity",
        "status": "manuscript and finite checks",
        "materials": "internal record"
      },
      {
        "topic": "Untwisted affine Lyndon periodicity",
        "status": "manuscript and finite checks",
        "public_url": "https://mantisdiscovery.github.io/affine-lyndon-periodicity/"
      },
      {
        "topic": "SL2(2^r) tensor-power spectrum",
        "status": "manuscript and finite checks",
        "public_url": "https://mantisdiscovery.github.io/sl2-tensor-spectrum/"
      },
      {
        "topic": "Nivat",
        "status": "full-scope proof manuscript; historical campaign archive",
        "external_peer_review_established": false,
        "full_formalization_established": false,
        "manuscript_url": "/research/nivat-manuscript.pdf",
        "manuscript_date": "2026-09-14",
        "manuscript_pages": 21,
        "manuscript_sha256": "6b3b5efdeb7aa549fdd0ca06cfda61b4be60241b369eb790d49b1f8fbf6b16e0",
        "historical_graph_url": "/nivat-history.json",
        "original_audits_transferred_to_this_revision": false,
        "historical_context": {
          "proposed_year": 1997,
          "manuscript_year": 2026,
          "years_between_proposal_and_manuscript": 29,
          "areas": [
            "two-dimensional symbolic dynamics",
            "combinatorics on words"
          ],
          "history_source": "https://drops.dagstuhl.de/entities/document/10.4230/LIPIcs.STACS.2020.14",
          "mathematical_context_source": "https://sites.math.northwestern.edu/brk762/papers/nivat.pdf"
        }
      }
    ],
    "scope": "Manuscripts, version-specific internal reviews, and finite computations with their recorded inputs and cases."
  },
  "method_url": "/evaluation-method.md",
  "architecture_url": "/technical-notes.html",
  "expanded_biology_workload": {
    "campaigns": 30,
    "agent_sessions_per_campaign": {
      "minimum": 37,
      "maximum": 44
    },
    "registered_agent_sessions": 1215,
    "completed_audit_outputs": 1215,
    "model": "huihui-qwen3.8-27b-abliterated",
    "simultaneous_model_call_limit": 4,
    "first_capture_at_utc": "2026-09-21T07:03:44.435000Z",
    "last_capture_at_utc": "2026-09-21T10:23:11.782000Z",
    "captures_with_live_model_call_telemetry": 24,
    "captures_without_live_model_call_telemetry": 6,
    "manifest_url": "/bio-expanded.json",
    "manifest_sha256": "bb34f05fec3e22a34c22a82c88290d4d6dd52df165b9ddd8d9e4d54cca05d5b2",
    "recording": {
      "url": "/media/biology-expanded.mp4",
      "sha256": "d17ae12b55a59eb19ddde37f9533b8629ad393504130afea7a3e68ffe30c6a32",
      "bytes": 712820,
      "duration_seconds": 35.262,
      "captions_url": "/media/biology-expanded.vtt",
      "provenance": {
        "re_encoded_date_utc": "2026-09-23",
        "note": "Re-encoded from re-rendered frames of the same 69 recorded native terminal frames (B05 capture, 2026-09-21T07:32:34.038Z to 07:33:09.299Z). Each frame's saved terminal cells were re-rendered with corrected glyph placement by the renderer used for the public stills, then encoded with the original per-frame timing, 1600x900 resolution and H.264 settings. No frame was added, removed or retimed, and the terminal text is unchanged. The caption line below the terminal reads 'up to 4 model calls at once' where the earlier encode read 'max 4 model calls generating'.",
        "previous_sha256": "ebd9cff29e5eb3568b07ab2c4c863bb6f93b2726de20b691ecd2c3b9834314a9",
        "previous_bytes": 589036
      }
    },
    "count_method": "Sum each unique campaign roster once. Completed outputs reconcile to distinct campaign/role final-response records; totals are separate from the original fleet workload.",
    "scope": "Thirty captured campaign batches performing source/design audits, sharing a four-call model queue. Session activity and request telemetry have separate timestamps; null identifies unsampled frame telemetry.",
    "archive_sha256": "e13735718c02717f86988cde48c06fb25e4bf3c4d0f67110a0a3dc267d6d6709",
    "preservation_checks": {
      "exact_transcript_outputs_verified": 1215,
      "all_bindings_stopped": true,
      "owned_writers_absent_before_and_after": true,
      "inputs_unchanged_during_archive": true
    },
    "request_accounting": {
      "calls": 1766,
      "completed": 1216,
      "failed": 550,
      "cancelled": 0,
      "rejectedWhileUnavailable": 102,
      "peak": 4,
      "maxActive": 4,
      "maxOutput": 500
    },
    "request_counter_scope": {
      "failed": "historical accepted-request failures remain unchanged, including queued requests stopped by the guard",
      "cancelled": "known client cancellations after transport diagnostics were added; earlier failures were not reclassified",
      "rejectedWhileUnavailable": "local requests rejected after the outage guard closed; no upstream request sent"
    },
    "assistant_message_stop_reasons": {
      "stop": 1215,
      "length": 1,
      "error": 308
    },
    "request_reconciliation": "1,216 successful requests produced 1,215 accepted final audits and one retained length-limited reply. Failed accepted requests include requests stopped in the local queue. Readiness probes are separate."
  },
  "benchmark_results": {
    "record": "/benchmarks.json",
    "proteinGym": {
      "metric": "Mean Spearman correlation using official protein/category aggregation",
      "methods": [
        {
          "name": "Reproduced Aster formula",
          "kind": "baseline",
          "full": 0.5235705908741475,
          "heldout": 0.5068562630588945
        },
        {
          "name": "Our research ensemble",
          "kind": "ensemble",
          "full": 0.5365378590465693,
          "heldout": 0.532041620615497
        }
      ],
      "models": [
        "ProSST-2048",
        "VenusREM",
        "GEMME"
      ],
      "heldoutGain": {
        "delta": 0.025185357556602405,
        "ci95": [
          0.011279338927029366,
          0.03769556123062094
        ],
        "bootstrap_draws": 9957
      },
      "coverage": {
        "assays": 217,
        "proteins": 186,
        "variants": 2465767,
        "heldoutProteins": 61,
        "heldoutAssays": 72,
        "developmentProteins": 125,
        "developmentAssays": 145
      },
      "replayStatus": "PASS",
      "missingPredictions": {
        "ProSST-2048": 0,
        "VenusREM": 0,
        "GEMME": 0
      },
      "method": "Select one of 63 equal-weight rank ensembles on development proteins, freeze it, then evaluate on held-out protein identities.",
      "scope": "Predictor comparison on a fixed UniProt-identity split. Full-benchmark scores include development data. The ensemble combines three models; the reproduced Aster formula uses ProSST.",
      "source": "Local completed experiment and separate evaluator replay",
      "comparisonRecord": "/research/proteingym-comparison.json",
      "replayRecord": "/research/proteingym-replay.json",
      "asterSource": "https://www.asterlab.ai/research/scaling_autonomous_research_to_thousands_of_agents",
      "benchmarkSource": "https://github.com/OATML-Markslab/ProteinGym",
      "sourceHashes": {
        "comparison.json": "a065b7a39ddaf318212abbbab93eec723f08ccd862d5c1cc66f4e2e4711bde98",
        "ensemble-independent-verification.json": "c425fba827794ef256abac575e7d129b0b508eea85c7aac5f98e0794974fdbfa",
        "ensemble-winner-frozen.json": "d9780dae2db64e28b391bc352432ae9015bd063513304ea80397ed8a5dc70348"
      }
    },
    "terminalBenchScience": {
      "version": "0.1",
      "metric": "Task resolution rate (%)",
      "results": [
        {
          "name": "GPT-6 Astra + Catenation",
          "score": 72.8,
          "origin": "Catenation"
        },
        {
          "name": "GPT-6 Astra",
          "score": 64.6,
          "origin": "Published comparison"
        },
        {
          "name": "Claude Fable 5.1",
          "score": 52.6,
          "origin": "Published comparison"
        },
        {
          "name": "Claude Opus 5",
          "score": 30.0,
          "origin": "Published comparison"
        }
      ],
      "gainPercentagePoints": 8.2,
      "publicComparisonSource": "https://openai.com/index/gpt-6-astra/",
      "benchmarkSource": "https://www.terminal-bench-science.ai/announcement",
      "source": "Catenation measurement with GPT-6 Astra; comparison scores are published results, checked against primary public sources.",
      "scope": "Catenation's task-resolution rate compared with the published score for each model. Differences are expressed in percentage points."
    }
  }
}