{
  "status": {
    "schema_version": 1,
    "project": "Pleiads",
    "state": "Active research",
    "current_experiment": "e006",
    "current_experiment_title": "E006 \u2014 Diversity Search",
    "checkpoint_date": "2026-10-04",
    "synchronized_at": "2026-10-04T13:26:23+00:00",
    "snapshot_kind": "frozen seed checkpoint",
    "provenance": "Laboratory-supplied frozen results; raw evidence files await separate ingestion.",
    "best_search_score": 5050,
    "realtime_confirmed": true,
    "replay_exact": true,
    "screenshot_matches": 1898,
    "screenshots_total": 1898,
    "e005_runs": 1980,
    "total_lab_runs": 2341,
    "research_target": 279100,
    "perception_model": "Qwen 3.5 9B / history3 / nearest4",
    "perception_exact": 90,
    "perception_opposite": 1.5
  },
  "experiments": {
    "e001": {
      "title": "E001 \u2014 Infrastructure & first gameplay",
      "state": "FROZEN",
      "branch": "QWEN / COMPUTER VISION",
      "evidence": "OBSERVED",
      "summary": "The first automated cohorts established observation, real game inputs, telemetry and deterministic replay."
    },
    "e002": {
      "title": "E002 \u2014 Perception benchmark",
      "state": "FROZEN v1",
      "branch": "QWEN",
      "evidence": "OBSERVED",
      "summary": "A balanced, independently labelled benchmark tests horizontal player localization under different visual conditions.",
      "frames": 200,
      "conditions": 24,
      "inferences": 4800,
      "results": [
        {
          "model": "Qwen 3.5 9B",
          "input": "4\u00d7 nearest-neighbour + 3-frame history",
          "exact": 90,
          "adjacent": 8.5,
          "opposite": 1.5,
          "median_latency_s": 1.984
        },
        {
          "model": "Qwen 3.5 27B",
          "input": "4\u00d7 nearest-neighbour + 3-frame history",
          "exact": 91,
          "adjacent": 7.5,
          "opposite": 1.5,
          "median_latency_s": 4.39
        },
        {
          "model": "Qwen 3.5 35B-A3B",
          "input": "4\u00d7 nearest-neighbour + 3-frame history",
          "exact": 86,
          "adjacent": 9,
          "opposite": 5,
          "median_latency_s": 4.031
        },
        {
          "model": "Qwen 3.5 9B",
          "input": "4\u00d7 nearest-neighbour, single frame",
          "exact": 80,
          "adjacent": null,
          "opposite": 12,
          "median_latency_s": 0.938
        }
      ]
    },
    "e003": {
      "title": "E003 \u2014 Controllers & Search",
      "state": "FROZEN",
      "branch": "COMPUTER VISION / POLICY SEARCH",
      "evidence": "OBSERVED",
      "summary": "Fast deterministic controllers and automated policy exploration form a separate branch from model-controlled gameplay."
    },
    "e004": {
      "title": "E004 \u2014 Qwen selection & timing",
      "state": "FROZEN",
      "branch": "QWEN",
      "evidence": "OBSERVED",
      "summary": "A higher-level Qwen policy selector was tested against deterministic controllers and Search.",
      "games": 30,
      "overall_max": 2020,
      "qwen_max": 1740
    },
    "e005": {
      "title": "E005 \u2014 Policy Search milestone",
      "state": "FROZEN",
      "branch": "POLICY SEARCH",
      "evidence": "REPRODUCED",
      "summary": "Automated policy search reached a real-time milestone with an exact deterministic replay and no model calls during the run.",
      "milestones": [
        3590,
        4450,
        5050
      ],
      "score": 5050,
      "survival_s": 183.24,
      "screenshot_matches": 1898,
      "screenshots_total": 1898,
      "model_calls": 0,
      "runs": 1980,
      "accelerated": 1974,
      "realtime": 6,
      "overnight_batches": 12,
      "ineffective_policies": 1052
    },
    "e006": {
      "title": "E006 \u2014 Diversity Search",
      "state": "ACTIVE at checkpoint",
      "branch": "POLICY SEARCH",
      "evidence": "OBSERVED",
      "summary": "Local timing optimisation plateaued. Diversity Search expands the behavioural search space.",
      "methods": [
        "Separate behavioural champions",
        "Local mutations",
        "Structural mutations",
        "Crossover",
        "New policy families",
        "Behavioural diversity",
        "Temporal score windows",
        "Effective-policy canonicalisation",
        "Exploration and exploitation branches"
      ],
      "results_available": false
    }
  },
  "timeline": [
    {
      "id": "event-001",
      "date": "2026-10-03",
      "order": 0,
      "title": "Lab infrastructure comes online",
      "description": "A Windows workstation with three RTX 3090 GPUs, 96 GB RAM and local Ollama connects to the MAME bridge. Input recording, restarts, video rendering and experiment logging are available.",
      "experiment": "e001",
      "badges": [
        "OBSERVED"
      ]
    },
    {
      "id": "event-002",
      "date": "2026-10-03",
      "order": 1,
      "title": "E001: the first automated cohorts",
      "description": "Vision, frame history, telemetry and episodic memory are exercised. Deterministic replay is demonstrated.",
      "experiment": "e001",
      "badges": [
        "QWEN",
        "REPLAY VERIFIED"
      ]
    },
    {
      "id": "event-003",
      "date": "2026-10-03",
      "order": 2,
      "title": "E002: a frozen perception benchmark",
      "description": "The 200-frame benchmark completes 4,800 inferences. Qwen 3.5 9B with nearest4 and history3 reaches 90% Exact and 1.5% opposite-side error.",
      "experiment": "e002",
      "badges": [
        "QWEN",
        "OBSERVED"
      ]
    },
    {
      "id": "event-004",
      "date": "2026-10-03",
      "order": 3,
      "title": "Search begins outperforming direct model control",
      "description": "Automated policy search explores controller policies. These results belong to Search, not to Qwen-controlled runs.",
      "experiment": "e003",
      "badges": [
        "POLICY SEARCH"
      ]
    },
    {
      "id": "event-005",
      "date": "2026-10-03",
      "order": 4,
      "title": "3,590 points in real time",
      "description": "Search reaches 3,590 in real time with exact replay.",
      "experiment": "e005",
      "badges": [
        "REAL TIME",
        "REPRODUCED"
      ]
    },
    {
      "id": "event-006",
      "date": "2026-10-03",
      "order": 5,
      "title": "4,450 points in real time",
      "description": "A further verified Search milestone reaches 4,450.",
      "experiment": "e005",
      "badges": [
        "REAL TIME",
        "POLICY SEARCH"
      ]
    },
    {
      "id": "event-007",
      "date": "2026-10-03",
      "order": 6,
      "title": "5,050 points \u2014 verified Search milestone",
      "description": "Real-time execution and exact replay: all 1,898 captured screenshots match. There are zero model calls during the run.",
      "experiment": "e005",
      "badges": [
        "REAL TIME",
        "REPLAY VERIFIED"
      ]
    },
    {
      "id": "event-008",
      "date": "2026-10-04",
      "order": 7,
      "title": "A plateau reveals ineffective switches",
      "description": "At the frozen checkpoint, 1,980 E005 runs and 12 overnight batches yield no score above 5,050. In 1,052 policies, planned switches occur after the final death.",
      "experiment": "e005",
      "badges": [
        "OBSERVED"
      ]
    },
    {
      "id": "event-009",
      "date": "2026-10-04",
      "order": 8,
      "title": "E006: Diversity Search begins",
      "description": "The method changes: broaden behaviours, canonicalise effective policies and balance exploration with exploitation. No new results are claimed.",
      "experiment": "e006",
      "badges": [
        "POLICY SEARCH",
        "ACTIVE"
      ]
    }
  ]
}