{
  "schema": "XIAOMAI_WEEKLY_PUBLIC_OBSERVATIONS_V1",
  "week": "2026-W43",
  "week_start": "2026-10-19",
  "week_end": "2026-10-25",
  "evidence_through": "2026-10-25",
  "record_updated": "2026-10-26",
  "language": "zh-Hant",
  "purpose": "本期已核對的研究與工程觀察。不同實驗、控制、來源讀回與封裝檢查分開判讀，不加總為整體通過率或能力分數。",
  "review_performed": "本次週報整理核對既有報告、正式結果與獨立簽核的指定內容及來源雜湊，沒有重跑模型，也沒有重開所有私人原始紀錄逐項複驗。",
  "evidence_availability": "完整核心、權重、私人保存點、原始張量、配置及完整日誌維持私有。公開摘要不能讓第三方獨立重現完整模型流程。來源 SHA-256 承諾只綁定位元組，不能單獨證明執行、正確性或能力。",
  "coverage_limits": "範圍限於本期已定位的 CPU 更新恢復、ledger 程式比較、三組合成程式解題、基本答題對照、外接研究助手、檢索修復及原生 9B 有界恢復與特徵紀錄，未宣稱完整盤點所有私人分支與工作。",
  "not_established": [
    "完整 9B／CUDA 與全 primitive 數值驗收",
    "正式舊 9B 保存點到新來源的遷移、冷恢復及下一步更新",
    "完整 decoder gradient 的首非有限值定位",
    "模型理解診斷並自主修正程式或研究方法",
    "一般自主研究、科學發現、持久能力提升或未見問題的因果效益",
    "六題以外的基本能力保留、另一獨立軟體環境或完整 True 能力",
    "實體斷電耐久、效能加速或市場獲利"
  ],
  "records": [
    {
      "id": "W43-01",
      "period": "2026-10-23",
      "observation_zh": "指定 CPU 小模型的實際更新成功、失敗及提交後觀測錯誤已接回下一輪研究輸入；寫入後故障回滾、同來源新程序恢復與重送防重複更新取得有界證據。指定觀測開關及控制組的 64 個事件、6,082 個 FP32 值、24,328 bytes 相同。這不是模型理解診斷、自主修正或完整 9B／CUDA 的證明。",
      "observation_en": "Actual update success, failure and post-commit observation errors reach the next research input on a bounded CPU small-model path. Declared post-write rollback, same-source fresh-process restore and duplicate-update prevention have bounded evidence. The specified observation-on/off and control runs match for 64 events, 6,082 FP32 values and 24,328 bytes. This does not establish diagnosis understanding, autonomous correction or full 9B/CUDA acceptance.",
      "metrics": {
        "observed_events": 64,
        "fp32_values": 6082,
        "compared_bytes": 24328
      },
      "limits": [
        "Declared boundaries rather than every primitive or storage location.",
        "Host-injected write faults rather than comprehensive natural-failure coverage.",
        "N3 remains PARTIAL; M3 remains PARTIAL_BLOCKED_TARGET.",
        "Windows saves retain COMMITTED_DURABILITY_UNCONFIRMED."
      ],
      "receipt_sha256_commitments": [
        "d2348e050ad6df6ece8639dffbeffc99a6ff8bab6085f8710ebc75bf2f522991"
      ]
    },
    {
      "id": "W43-02",
      "period": "2026-10-22",
      "observation_zh": "八個新合成世界、三組，共 24 場、96 回合。FULL／關跨回合記憶 M0／不提交更新 W0 的平均留出正確率為 2.08%／5.99%／2.08%，完整組未通過兩項預登記優勢門檻。正式更新全部為零，因此本批沒有量到成功更新的效益，也不能據此刪除記憶或學習機制。原無效批次、生成失敗及考後驗證器勘誤分別保留。",
      "observation_en": "Eight new synthetic worlds and three arms produced 24 sessions and 96 episodes. FULL, no-cross-episode-memory M0 and no-committed-update W0 had mean held-out accuracy of 2.08%, 5.99% and 2.08%. FULL failed both preregistered advantage criteria. Every formal committed-update count was zero, so this batch did not measure successful-update benefits or justify removing memory or learning. The prior invalid batch, generation failures and post-run checker corrections remain distinct records.",
      "metrics": {
        "worlds": 8,
        "arms": 3,
        "sessions": 24,
        "episodes": 96,
        "heldout_episodes": 48,
        "mean_heldout_accuracy": {
          "FULL": 0.020833333333333332,
          "M0": 0.059895833333333336,
          "W0": 0.020833333333333332
        },
        "formal_committed_updates": {
          "FULL": 0,
          "M0": 0,
          "W0": 0
        },
        "preregistered_full_advantage": {
          "over_M0": false,
          "over_W0": false
        }
      },
      "limits": [
        "Bounded synthetic coding tasks rather than a general-autonomy or scientific-discovery test.",
        "The complete original native research loop was not exercised.",
        "No fully correct valid training proposal opened the existing update gate.",
        "Separate public update controls are excluded from formal scores.",
        "Examiner and verifier were different agents from the same model family, not external third parties."
      ],
      "receipt_sha256_commitments": [
        "08224b54b519fcfc3189d2109099dbd817a912763777092f73bc35a6c8677835"
      ]
    },
    {
      "id": "W43-03",
      "period": "2026-10-23",
      "observation_zh": "同一份 0.5B 模型在不掛輸出頭、初始輸出頭、公開正控兩次更新後輸出頭三種狀態，六題的生成 token 與回答相同。基線已錯兩題算術；程式題三組皆達 64 token 上限。受測輸出頭在這六題沒有新增可見退化，不能推廣為完整小麥、9B 或 True 的能力保留。",
      "observation_en": "The same 0.5B model generated identical tokens and answers on six prompts with no output adapter, the initial adapter and an adapter after two public control updates. The baseline already failed two arithmetic prompts; the coding prompt hit the 64-token cap in all three states. No additional visible answer regression was observed on these prompts. This does not establish preserved capability in full Xiaomai, 9B or True.",
      "metrics": {
        "prompts": 6,
        "states": 3,
        "max_new_tokens_per_prompt": 64,
        "baseline_arithmetic_errors": 2,
        "coding_token_cap_failures": 3,
        "public_control_updates_before_third_state": 2
      },
      "conditions": [
        "Qwen2.5-Coder-0.5B-Instruct",
        "CPU FP32",
        "greedy decoding",
        "three fresh processes; one run per state"
      ],
      "limits": [
        "The updated adapter changed logits while generated tokens on these prompts stayed identical.",
        "No training or full workbench was run during this answer check.",
        "No second independent software environment was tested.",
        "Aggregate checker counts are not a capability or pure-format pass rate."
      ],
      "receipt_sha256_commitments": [
        "fc8c1333c830813bd2f7c4d2aad6ab9a4ae8f312d162b0f3cd45be3fcf83c08d"
      ]
    },
    {
      "id": "W43-04",
      "period": "2026-10-23",
      "observation_zh": "CodeGraph CLI 與 MCP 查詢在固定來源上實際使用，協助工程作者查找更新、觀測與恢復接線。動態呼叫缺邊與同名錯邊仍存在；工具索引不是模型執行或自我理解的證據，也未量測效率改善。",
      "observation_en": "CodeGraph CLI and MCP queries were used on fixed sources to help engineering authors locate update, observation and restore connections. Missing dynamic-call edges and same-name false edges remain. Indexing is not model-execution or self-understanding evidence, and no efficiency improvement was measured.",
      "receipt_sha256_commitments": [
        "78586255e2a9ce63851fa6eee959ac46b6e62b5516c1f8621076d1a15cb986b5"
      ]
    },
    {
      "id": "W43-05",
      "period": "2026-10-19/2026-10-25",
      "observation_zh": "前期多層 FP32 0/14 通過與負控制 104/112 保留原結論，本週不同契約的通過沒有消除那些失敗。接續工程需補原生新保存物來源清單及冷載入，再推進完整研究更新；能力實驗另行預登記，用未見實例比較原方法、系統修訂與無實質修訂的對照。",
      "observation_en": "Prior multilayer FP32 0/14 passes and negative controls 104/112 retain their original conclusions. Passes under this week's different contracts do not erase them. Next work must resolve the native new save's source inventory and cold loading before full research updates. Capability experiments require separate preregistration and unseen-instance controls.",
      "historical_public_source": "https://xiaomailab.com/updates/2026-W42/",
      "receipt_sha256_commitments": []
    },
    {
      "id": "W43-06",
      "period": "2026-10-19/2026-10-21",
      "observation_zh": "10/19 正式保存續跑與 SQL 提交後熱／冷重試取得有界證據，仍依賴外部 journal 及原路徑 episode。10/21 將執行完成和任務答對分開評分，去除指定等價重工，修復同碼 marshal bytes 不穩定造成的冷恢復指紋假拒；新 tiny 完整第二研究回合與冷續跑的指定持久狀態一致。不代表任意 Python 狀態、所有 CPU 工作或能力驗收。",
      "observation_en": "Oct 19 formal save/continuation and SQL post-commit hot/cold retries have bounded evidence, with external journal and original-path episode dependencies retained. Oct 21 separated completed execution from correct answers, removed specified equivalent rework and repaired false cold-restore rejection from unstable marshal bytes. A tiny complete second research round matched declared persistent state after cold continuation. This is not arbitrary-Python, all-CPU or capability acceptance.",
      "receipt_sha256_commitments": [
        "406d753f86e42f26b5f4093ba735bb8520a88f03e0fa8e0f9032eb0a0393232a",
        "4f5937b581003dc445de4f72f8c1575ac87a53e578c8f8b2b483ed6498a70ce3"
      ]
    },
    {
      "id": "W43-07",
      "period": "2026-10-20",
      "observation_zh": "ledger v5／v6 是作者改版的受測程式，在固定作者手寫 H／P 表示語言內比較。v5 280 場完成；v6 新實例 280 場中 279 完成、1 逾時，六條預登記預測四成立，Q3 時鐘與 Q6 時限失敗。一步狀態、延遲輸入與兩步前輸出在受測預算內全作答全對；計數全 null，週期 7 已作答錯誤約半，含 validated=true 仍錯。未驗新表示種類、主迴圈整合、保存重啟或持久轉移。",
      "observation_en": "Ledger v5/v6 are author-revised tested programs compared within the author's fixed H/P representation language. v5 completed 280 sessions; v6 completed 279 of 280 fresh-instance sessions with one timeout. Four of six preregistered predictions held; Q3 clock and Q6 time-limit gates failed. Tested state/delay/output-memory cases were fully answered and correct; counting always abstained, while period-seven answered errors were near half, including validated=true cases. New representation types, main-loop integration, save/restart and persistent transfer remain untested.",
      "metrics": {
        "v5_completed": 280,
        "v6_completed": 279,
        "v6_timeouts": 1,
        "v6_sessions": 280,
        "preregistered_predictions_met": 4,
        "preregistered_predictions_total": 6
      },
      "receipt_sha256_commitments": [
        "4f38fb331dcad919a5e09fb084c2d03963c947180fe6fb0a9064ca953bb49560",
        "e30dcf5e42585ece6deddc160b0952efac55ebce424af81e20d28fc8eb31022f",
        "e00a38e2d92460abd5e9791950eca2dd1c509afe6d5c1153b42fe69b0a02281e"
      ]
    },
    {
      "id": "W43-08",
      "period": "2026-10-25",
      "observation_zh": "真 Ornith-1.5-9B-Q4_K_M、llama.cpp 與 Instructor 接上本機唯讀研究助手。ECC 初輪分輪／單輪完整正確同為 1/4，分輪耗時與 token 更多。修復版第一個完成四題成對區塊，完整正確舊 0/4、新 1/4，完整事實 1/12→7/12，必要引用 4/5→5/5；成本也增加。81 個額外或中止案例另留，並非事前預註冊八案例、未見題或單因素因果比較。外部輸出不執行程式、不採納方法、不呼叫原生更新。",
      "observation_en": "Real Ornith-1.5-9B-Q4_K_M, llama.cpp and Instructor powered a local read-only research assistant. Initial iterative and single-round retrieval both scored 1/4 fully correct, with higher iterative time/token cost. The repair's first completed four-task paired block scored old 0/4 versus new 1/4 fully correct, complete facts 1/12 to 7/12 and required citations 4/5 to 5/5, at higher cost. Another 81 extra/interrupted cases remain separate. This was not a preregistered eight-case, unseen-task or single-factor causal experiment. External output does not execute code, adopt methods or invoke native updates.",
      "metrics": {
        "paired_tasks": 4,
        "complete_correct": {
          "old": 0,
          "new": 1,
          "denominator": 4
        },
        "complete_facts": {
          "old": 1,
          "new": 7,
          "denominator": 12
        },
        "required_source_coverage": {
          "old": 4,
          "new": 5,
          "denominator": 5
        },
        "workflow_seconds": {
          "old": 211.703,
          "new": 363.577
        },
        "reported_tokens": {
          "old": 12194,
          "new": 26626
        },
        "calls_per_arm": 11,
        "completed_responses_with_stop": 22,
        "extra_or_interrupted_cases_separate": 81
      },
      "limits": [
        "Costs include decisions and final answers for the selected complete block, not all development work.",
        "The original larger repeated plan was reduced; block selection was fixed before anonymous scoring.",
        "Context 8192, final-token limit 768, decision-token limit 192, same model and temperature 0/seed 42.",
        "Source strategy, prompts and schema changed together; no single-repair causal attribution.",
        "Some interrupted cases have unknown final usage."
      ],
      "receipt_sha256_commitments": [
        "f88654db7e757ff9e32e58fd3e8f9b42d21fb6825e398cdb1ac85a390fe84a9f",
        "af6dd92e391da5f13a9c1a226b57c2464a119912ccb69f73fc77e708be9df43b",
        "1ca54bd07ab84e02de3e0d6b9b4a83a78a06eb239b95f694621a3184740bf9a6",
        "83c1a0938e3e343c50051068c5982bc1eee92f7cff3996eb3668c6d20b9c9073"
      ]
    },
    {
      "id": "W43-09",
      "period": "2026-10-25",
      "observation_zh": "首輪原生 9B 成功映射原 BF16 底座，ancestor 指紋階段因 RAM guard 中止，自訂保存點尚未還原。後續分塊 digest overlay 實載原 9B 與較舊 CPU 自訂狀態，22 類狀態相同、41 records 保留，另存未更新狀態。研究與更新未到達，新保存物尚未 cold-load。來源清單舊 115、新 104，strict audit 的 only_expected_source_migrated=false 保留；不是完整來源閉包或最新 CUDA 保存點遷移驗收。",
      "observation_en": "The first native 9B attempt mapped the original BF16 base but hit a RAM guard during ancestor fingerprinting before custom restore. A later streaming-digest overlay loaded the original 9B and an older CPU custom state, matching 22 state categories and retaining 41 records, then saved unchanged state. Research/updates were not reached and the new save has not been cold-loaded. Source inventory changed from 115 to 104; strict only_expected_source_migrated=false remains. This is not complete source closure or migration acceptance for the latest CUDA checkpoint.",
      "metrics": {
        "matched_state_categories": 22,
        "retained_research_records": 41,
        "old_source_inventory": 115,
        "new_source_inventory": 104
      },
      "status": "PARTIAL_RESTORE_SAVE_ONLY",
      "research_executed": false,
      "parameters_updated": false,
      "new_save_cold_loaded": false,
      "only_expected_source_migrated": false,
      "receipt_sha256_commitments": [
        "f49ed5c2414e28338ea1189e64599ddf9459f1ac08877165b424980c15155ba3",
        "e3d76e234c525fa2b46e4c3de9b645c7076ee0541d61b730aca0c6991e7bcec5"
      ]
    },
    {
      "id": "W43-10",
      "period": "2026-10-25",
      "observation_zh": "外接量化推論和原生特徵比較分開判讀。外接同請求／同輸出 bytes 的 128 token A→B→A 短對照，實際 offload 28/34→34/34→28/34，輸出吞吐率 12.298／31.547／11.693 token/s；只有一次 B，背景負載不同，不是穩定端到端加速。原生已恢復 9B 的四 token 前綴只比 1×32 FP32 特徵，CPU 24.641 秒、GPU 30.734 秒；容差通過而 GPU 本次較慢，非生成速度。",
      "observation_en": "External quantized inference and native feature comparisons remain separate. An identical-request/output 128-token A/B/A sample actually offloaded 28/34, 34/34 and 28/34 layers and reported 12.298, 31.547 and 11.693 output tokens/s. With one B sample and different background load, it is not stable end-to-end acceleration. The restored native 9B four-token prefix compared only 1x32 FP32 features: CPU 24.641 seconds, GPU 30.734 seconds. Numerical tolerances passed while GPU took longer in this pair; this is not generation speed.",
      "limits": [
        "External length termination was planned at 128 tokens, not a completed research answer.",
        "Native comparison was one CPU-to-GPU pair with two threads, not alternating repeat statistics.",
        "Native relative RMSE 0.0021043773740530014 and maximum absolute difference 0.004306793212890625 met 0.005/0.25 gates; not bitwise equality.",
        "Native per-matrix weight transfers are observed; copy and compute were not independently profiled.",
        "Neither measurement establishes complete Xiaomai research/update throughput."
      ],
      "receipt_sha256_commitments": [
        "a4a5c9bab17d9d06763f69a045215cc97daebb20da1ce81e09b4fc149d2945a4",
        "aad8edf49601f95497b3be52e3f8d2332c8c7d3eefbb6eaa7c0930961bc4701e"
      ]
    }
  ],
  "open_source_tools": [
    {
      "name": "CodeGraph",
      "repository": "https://github.com/colbymchenry/codegraph",
      "version": "1.6.2",
      "source_ref": "6560052a6f856855d3f71eee838fd66ccfa4285d",
      "license": "MIT",
      "license_url": "https://github.com/colbymchenry/codegraph/blob/6560052a6f856855d3f71eee838fd66ccfa4285d/LICENSE",
      "role": "executed_tool",
      "usage_zh": "實際使用 CLI 與 MCP 建立來源索引、定位符號並查更新、觀測和恢復接線；圖譜仍有缺邊及錯邊，需回原碼核對。",
      "usage_en": "Used the CLI and MCP interface to index sources, locate symbols and trace update, observation and restore connections. Missing and false edges still require source checks."
    },
    {
      "name": "Instructor",
      "repository": "https://github.com/567-labs/instructor",
      "version": "1.17.0",
      "source_ref": null,
      "license": "MIT",
      "license_url": "https://github.com/567-labs/instructor/blob/main/LICENSE",
      "role": "executed_tool",
      "usage_zh": "將本機 Ornith 回應接成結構化研究審閱及檢索決策；格式與語意分開核對，不授予執行或 learner 更新權限。",
      "usage_en": "Produced structured research reviews and retrieval decisions from local Ornith responses. Format and meaning were checked separately; no execution or learner-update authority was granted."
    },
    {
      "name": "llama.cpp",
      "repository": "https://github.com/ggml-org/llama.cpp",
      "version": "b11669",
      "source_ref": "620d2225cc436ab0b001e3200e590e3b286ef9ab",
      "license": "MIT",
      "license_url": "https://github.com/ggml-org/llama.cpp/blob/620d2225cc436ab0b001e3200e590e3b286ef9ab/LICENSE",
      "role": "executed_runtime",
      "usage_zh": "以 llama-server 載入 Ornith GGUF，提供 CPU／Vulkan 的本機 HTTP 推論；服務外接量化助手，未移植小麥原生 learner 或自訂保存點。",
      "usage_en": "Loaded Ornith GGUF with llama-server for local CPU/Vulkan HTTP inference. This served the external quantized assistant, without porting the native Xiaomai learner or custom checkpoints."
    },
    {
      "name": "ECC",
      "repository": "https://github.com/affaan-m/ECC",
      "version": "ecc-universal 2.2.3",
      "source_ref": "90606c8315ee09cafc303515c5f33024db6b9882",
      "license": "MIT",
      "license_url": "https://github.com/affaan-m/ECC/blob/90606c8315ee09cafc303515c5f33024db6b9882/LICENSE",
      "role": "workflow_reference_locally_adapted",
      "usage_zh": "參考 iterative-retrieval 的分輪檢索、缺口檢查和補讀流程，由本地實作接上 CodeGraph 與 Instructor。沒有整套安裝 ECC 或啟用全域 hooks，也不是參數學習。",
      "usage_en": "Its iterative-retrieval workflow informed the local staged-retrieval, gap-checking and follow-up implementation through CodeGraph and Instructor. The complete ECC package and global hooks were not installed; this is not parameter learning."
    },
    {
      "name": "Pydantic",
      "repository": "https://github.com/pydantic/pydantic",
      "version": "2.14.1",
      "source_ref": null,
      "license": "MIT",
      "license_url": "https://github.com/pydantic/pydantic/blob/main/LICENSE",
      "role": "schema_dependency",
      "usage_zh": "供 Instructor 接線的嚴格型別與欄位檢查；結構通過不等於內容正確。",
      "usage_en": "Provided strict type and field validation for the Instructor integration. A valid structure does not establish correct content."
    },
    {
      "name": "OpenAI Python SDK",
      "repository": "https://github.com/openai/openai-python",
      "version": "3.3.0",
      "source_ref": null,
      "license": "Apache-2.0",
      "license_url": "https://github.com/openai/openai-python/blob/main/LICENSE",
      "role": "local_http_client_dependency",
      "usage_zh": "經 Instructor 呼叫本機 OpenAI 相容 HTTP 端點；不表示這批實驗使用 OpenAI 雲端模型。",
      "usage_en": "Called the local OpenAI-compatible HTTP endpoint through Instructor. This does not mean these experiments used an OpenAI cloud model."
    }
  ],
  "tool_attribution_checked_on": "2026-10-26",
  "tool_attribution_scope": "Major research tools, locally adapted workflow reference and directly verified integration dependencies; not a complete software bill of materials. Official upstream license files were checked; no additional model execution."
}