{ "schema": "XIAOMAI_WEEKLY_PUBLIC_OBSERVATIONS_V1", "week": "2026-W42", "week_start": "2026-10-12", "week_end": "2026-10-18", "record_updated": "2026-10-18", "language": "zh-Hant", "purpose": "本期已核對來源的工作索引；同一研究線的候選、修補、測試與保管分開列項，不是獨立成果或通過數的加總。", "review_performed": "本次整理讀取歷史收據、結果及來源並核對雜湊，沒有重新執行模型。部分觀察沿用既有結果審查，未於本次重開大型封存逐件複驗。", "evidence_availability": "完整核心、權重、私人保存點與原始張量維持私有。此處只公開已審閱的摘要與來源雜湊承諾。雜湊不能單獨證明執行、正確性、能力或完整可重放。", "coverage_limits": "依本週已定位的研究與工程來源整理；沒有宣稱所有私人分支、對話或未留存工作都已完整盤點。", "not_established": [ "全參數訓練與一般數值等價", "任意自由輸入與長輸入完整驗收", "跨實體機器與完整封裝重啟", "自主研究、因果能力提升或效能加速" ], "records": [ { "id": "W42-01", "period": "2026-10-13", "observation_zh": "研究流程周邊也補了方法稽核、原始紀錄會計、只啟動一次和子程序結束控制，先用人工反例確認失敗會真的被擋下來。這些是在讓後續實驗的紀錄更可信，沒有拿控制測試代替模型能力。", "receipt_sha256_commitments": [ "25128afe0ceda1713b8db7187d06c8ea47d9e43e5d4a84e18e44895009e53b9d", "31463f398769a3f8ae4bcdec19db6c74ffe30ecc29f172ac96e6b44427454277", "4f9a32ce796aa93dabcaf6b66ef889922684b1e1caa3fdb0643e534e71760abb", "4fddbdd99fe66d4b01ccca2cc727ee554327878aa65f574a12c82e856caf1979", "74bfbcffdbd717c4f0a1ee4a24c0dc79b240b8f79cb515fcd5b53da2cb5f66b9", "9037bb193ad75a25dc65ffebddafc59fdebfe1f3f9cc3bc7a53d611a56f205b3", "e9da2acd526af5fbc1885ea7433f617cc10f672a4cb1f4863661e1de3d2a103b" ] }, { "id": "W42-02", "period": "2026-10-13", "observation_zh": "MLP 的 BF16 版本也有另外做小型 CPU 診斷，確認每一步的取整和最後才轉回 BF16 並不一樣。探索候選有通過這些人工案例，但原 A100 環境、真尺寸和原生 BF16 仍要另外驗。", "receipt_sha256_commitments": [ "46274d87ea857d1253d9234086e4654d7c3c8fc6e2e9c315bd13a462193cebf9", "e46756a2882ee2e2416af09ff7e7c77e921dfa42e930e6835331c7a69d475592" ] }, { "id": "W42-03", "period": "2026-10-13", "observation_zh": "這週也把已結束的研究結果連同失敗原件保存、讀回核對，再結束原來的 GPU 實例。保存完成能證明資料有留下來，不能代替模型冷恢復或能力驗收。", "receipt_sha256_commitments": [ "46024df46a9b4d28d549eca860a773df7107b06d9913c570474a383f42e2fd0b", "8e8978187d32126a2d9beb8b69e53e14417584d1c92a67f91e8a3808a9e15aaa", "c92f00253836380dd6f55ee59fd22fe72934f0c11c3be1d359d677cfd8592140" ] }, { "id": "W42-04", "period": "2026-10-13", "observation_zh": "原研究更新留下的梯度和狀態資料，也補做了來源綁定的逐位元核對。新的只讀核對有一致，但原執行器的拒絕紀錄還是保留，不能把它改成整段部署通過。", "receipt_sha256_commitments": [ "554d2c3e8bfdd35cdb36458b04ba6a2a56efb00c2221f50f6f63e2cfc444e13b", "bf658fabb61cb309d1e4b436858d87c28d5d8e3eafe5530c24d3f0ddf4c056eb" ] }, { "id": "W42-05", "period": "2026-10-13", "observation_zh": "方法保存和重新載入也查到一個細節：換行格式被讀檔方式改掉，會讓原件核對失敗。修補後用人工方法跑過保存、關閉原程序、新程序執行和外部評分，但這還不是模型自己發明方法的成果。", "receipt_sha256_commitments": [ "024b9adb7d3feab9f43e48c1e9d96fb9c22aeb921dc035cd7442cfc53121bd1a", "7fd1fe4f999cd9708936a63c096ceee6f2b8b73494d443ca93a6738f907a66de", "df9155cf898244bf79b499900c779fe744200abf01dfc1361639ce54986f04fc" ] }, { "id": "W42-06", "period": "2026-10-13", "observation_zh": "10 月 13 日先在 GPU 上核對原模型的一個運算區塊，指定的前向與反向比較通過固定門檻。使用的是真權重轉成 FP32，加上人工輸入，範圍仍只有這個區塊。", "receipt_sha256_commitments": [ "07e526f0671f7c541a71b1a0bbf9c69ccc469ecb047bb277d33a787e20b29b07", "f79909f4022e525bb378498fbe229111cac46a468d95a08797b6e73e1659717a" ] }, { "id": "W42-07", "period": "2026-10-13", "observation_zh": "真模型的文字路徑有核對到和原採樣結果一致，不過輸出內容仍有格式錯誤。研究方法那條線也在原本允許的修正次數用完後被拒絕，還沒走到方法執行和新任務評估。", "receipt_sha256_commitments": [ "2a6ec416bb580e696b66061fec13dfa0dfc092ebc27afd91cacc1688d2fb7f90", "74bfbcffdbd717c4f0a1ee4a24c0dc79b240b8f79cb515fcd5b53da2cb5f66b9", "facda17fefe13e85e9c38db5ef940e4c78c6a441fa764541a30dd48947ee889b" ] }, { "id": "W42-08", "period": "2026-10-13", "observation_zh": "數值核心也先用封存的實際輸出資料，核對幾種基本運算、明確反向和操作紀錄。它能補出這一小段運算與拒絕反例的證據，還不能代替完整研究梯度或整個模型的追蹤。", "receipt_sha256_commitments": [ "5a36df9a79fe895ce06fe24888cdaf03d1a642094cd272f012230970341d31ba", "5e93b79e1d20748c77057d6d388fd296b8f458992a7dc7f978bf935ff477cb7e", "5ea2001dfb9c7346a771dc3444e218a6f4733293715def786391f87f73925800" ] }, { "id": "W42-09", "period": "2026-10-13", "observation_zh": "同一份實際輸入資料，也另外拿來核對 17 組研究端梯度，指定比較通過。這補的是這一小段計算的證據，還沒有證明完整模型反向或保存後接續。", "receipt_sha256_commitments": [ "29f011ebbb35c2ec3dd0ddfa1e0d10df68014ef332305262ea9f966ea702e647", "52428337ba843814efbe4bf7035511bbfcc4554981131e974c71a1f748f2554d", "61f71c1afce91ef661b0ff26b5b90d513ab9cadce47ce338e71370b16c80a6ca" ] }, { "id": "W42-10", "period": "2026-10-14", "observation_zh": "10 月 14 日先把注意力、GDN 和正規化等區塊的前向與手動反向寫成可檢查的 Python，並用 CPU 人工案例逐項對照。BF16 也有探索結果，但還不能當成真權重或 GPU 的正式驗收。", "receipt_sha256_commitments": [ "187a1eeb86a500cd66d30c6f58e2ae16b9c6c2fb64aaf9422eb364841499c378", "1b13db195a9fde2cdd3eb1499aedaf9c325ec95039ed67a90d4cd7968d8a9183", "a737cb51889538dcb55f714444ddb5d735f1bac6560fac0e750b51be77727ca3", "b986f1dbde3eeab064c5d9720334c405b74d44ac24e9417465d95931aad63dd3", "fda6725ad01b11c2c032bc6aff0a53fbe053846221428360f4c6bd2c6be5138c" ] }, { "id": "W42-11", "period": "2026-10-14", "observation_zh": "我也回頭拆解原本生成很慢的紀錄，確認正式解碼已有快取，並補上後續可以量階段耗時的工具。這次只是分析與工具核對，還沒有量到可宣稱的加速。", "receipt_sha256_commitments": [ "83e8f6d5150235821af1eb7e9e8a3692f8cdf747188bf5b7379ec96ceafb74da", "a737cb51889538dcb55f714444ddb5d735f1bac6560fac0e750b51be77727ca3" ] }, { "id": "W42-12", "period": "2026-10-14", "observation_zh": "把區塊接成多層後，14 個新種子 CPU 案例仍全部沒有通過原 FP32 門檻。重算版本的位元一致性和部分控制雖然有通過，完整協定仍然是失敗。", "receipt_sha256_commitments": [ "8e99a250afc527a1dc94a5ee6176d6ecb3d4672bd9db866f264e0524de3f24fd", "90415f251c491dfed68cb050ee8dde328b5110a4e7a5d730911b0f1f83048393", "df8d9d3bdb78d82ca0e049ffea08d8b3849a70dddd097ab9b3bd54f26ec819a6" ] }, { "id": "W42-13", "period": "2026-10-14", "observation_zh": "負控制本身也查到非連續資料可能只改到副本的問題，所以修補另外保存，先用人工反例確認有改到真正資料。原凍結版本與四個案例的失敗都保留，這六項檢查沒有被當成新的模型驗收。", "receipt_sha256_commitments": [ "290a2a80a588ffb715ba1963f973c96da5aad9f8858b765ffaf00bf33da8ca4d", "df8d9d3bdb78d82ca0e049ffea08d8b3849a70dddd097ab9b3bd54f26ec819a6" ] }, { "id": "W42-14", "period": "2026-10-14", "observation_zh": "同一天也把這輪來源和原始成功、失敗紀錄一起保存到私有版本庫，讓後面可以對回同一批原件。資料交付通過和科學驗收失敗分開記錄，沒有用發布完成把失敗蓋掉。", "receipt_sha256_commitments": [ "4c24dcc273132c81834a892f6bb33289e7a2029372bb79d0838206888ee26152", "8eaf95cf762a759aabddfd974634fc46605a37af16edb28c5bf052f9acba9a50" ] }, { "id": "W42-15", "period": "2026-10-15", "observation_zh": "接著回到原部署實際使用的 GDN 和正規化版本，補齊 Python 前向與明確反向的 CPU 對照。也有用封存的真實輸入核對，但當時的上游梯度沒有留下來，所以這部分仍搭配人工梯度，不能叫作原訓練重播。", "receipt_sha256_commitments": [ "038f10adec5bf53d0f1f41611043c4a99df9d3b407428a982cd605e556f6afef", "40c12b59de3b724d09b98524a5f620d5e882cc7a9ff3a8b2d7fd1c22038706d1", "722b0fe99488800ea7b7d17866dec8978a5fef7d25f7c2ccb0badf7d012975cf", "80e81cb9591a8841889ccbcb5d476493c3218c978ed85087d6b467a8fdca7054", "bcf99fa10153d41b8d3617edec558d802a4199738915bcb35de26f695cde6bd1", "d5f3a8a6c275d156b2673c248bb98330e3b9d39f83d6ba97ed0bbbef0a8ff6d7" ] }, { "id": "W42-16", "period": "2026-10-15", "observation_zh": "保存也從組件資料往前接到原本的整體 graph，文字更新和研究信用交易都有換新程序再走下一步的紀錄。Group 和 Decision 也分別做了冷啟核對，各條支持的設定還是分開記錄。", "receipt_sha256_commitments": [ "04d8104442dc2235c89b049ec3c12ec5b1abaea487285f81c27fc3365ee8d8f3", "89e90c8a668e284a0511705539b7d837815aa01fa710dcb97e6610dd50b49803", "8f434be5b71fd0d4b586572418802d2f5f61525ff4f3487177c2f0cab8f24e40", "c561c4f92f359ff2e6e23752678663ecbbdfbb3f6566c5284f2686b07f62c70e" ] }, { "id": "W42-17", "period": "2026-10-15", "observation_zh": "後面把原本的文字學習、研究信用、群組策略和決策更新入口逐項接回去，讓原方法實際走過新的 Python 計算與更新。這幾條有各自的 CPU 交易和反例證據，還不能把不同測試拼成完整 9B 已接管。", "receipt_sha256_commitments": [ "169481c2e3ff3bfa29d31ae4ff521d7db8d08bd1780c83e53262cbcb007e744a", "24454bc2a39a6f9fdbc8086a0fc4e3476d2310c9bc293ccd33017a169b66b218", "724dea24be3526f0f87f27d15ecc6e3419eea9646ac7d58e8faebf7f36db3cb2", "89e90c8a668e284a0511705539b7d837815aa01fa710dcb97e6610dd50b49803", "8f434be5b71fd0d4b586572418802d2f5f61525ff4f3487177c2f0cab8f24e40", "a27e8d470604db922a56c20514fb99d571cc21b3a57759199d7034fc84355ae4", "e8427465ce25259a1e2cbbad671300a6777ee4864b2aedc388285e9ac6de47fd" ] }, { "id": "W42-18", "period": "2026-10-15", "observation_zh": "數值偏差也有繼續查，補償候選讓超界座標變少，但 14 個案例仍全部失敗，也更慢，所以沒有換進主流程。另外查到部分座標的兩個參考門檻根本不相交，需要先把數學目標和原框架相容性分開說清楚。", "receipt_sha256_commitments": [ "28154053e1451d46e5d42f86647ecf23c9185d2f105f84a8df55b71f4ea7605a", "f309bb8abc510bb20b080487708c20014bcc61cb58dca02da041e008c5219270", "f36f2a2dd81439b8c31391ab2c32df43f4be08111a21e58fe41ae6e1e1fa42be" ] }, { "id": "W42-19", "period": "2026-10-15", "observation_zh": "10 月 15 日先用縮小模型把前向、手動反向、更新、保存和新程序的下一步接成同一條流程。記憶模組和原更新名單也一起接入，能更新的和凍結的角色分開核對；這仍是 CPU 小模型。", "receipt_sha256_commitments": [ "0f64ac8e46ad105e1061231746e432cb72bf02366d9d2cf7f75c8bc8d39527fd", "3d3a8ef47e2bb2a5f2093ea9c8e792142b6fd451a27ba248f79cbda07f5cf5ed", "ad81b8126bc99422426abfaad6e634fab1021024ba330484b0a4f039b5950e01" ] }, { "id": "W42-20", "period": "2026-10-15", "observation_zh": "這週也補了張量操作紀錄和研究循環的 Python 接頭，讓過程留下可檢查的資料。紀錄工具有做開關和損壞反例核對，但研究循環這條仍是工程測試，沒有用它宣稱模型已能自主研究。", "receipt_sha256_commitments": [ "00eceb0cf5426404207273ebe02edfcd08b3fa6a373e15466268b89e32405956", "a2aaddea8d64da45d54ee17291020263663cedb4c0dd8f031bf4964ad1fccfc6", "aafbd06f007af7bb81f73d9616f902aa664c274c36ee7c8a1c0fa4e5e65d3f59" ] }, { "id": "W42-21", "period": "2026-10-15", "observation_zh": "到這一段，原本的啟動和 constructor 也確實接進同一個小模型流程，研究信用、Group、Decision、保存和新程序下一步一起核對。前面冷載找出的來源 owner 缺口也保留失敗，再用新版本補驗；完整 9B 載入和模型生成仍沒有在這輪完成。", "receipt_sha256_commitments": [ "0eb1e63c1e83ea7ac7cd7367b5980c8afb003de8732b61e09bb56b129f58db08", "8f0044e5fce74366a0711a025e1704de5701f3bac4bc520b750bb60694571ec7", "a76a874bada2291e9436d8c7738d46a7ca7039c0bb20a1741569983bcf33c842" ] }, { "id": "W42-22", "period": "2026-10-15", "observation_zh": "接線後也繼續處理較大的更新狀態，以及 BF16 儲存、FP32 計算和 CPU master 的搬移。這些支持域都有分開實跑，不過完整模型的梯度、activation 和交易備份還沒有因此變成全程串流。", "receipt_sha256_commitments": [ "41d0c05f05b6746e21cd6e79a2742aea452de79383bf5e5aec2a57338b013760", "7de1d5a4f9ee1e441ba02a287f66f1c3efd10f3a74a2ceb7c6f1271e11dda2ab", "8f0044e5fce74366a0711a025e1704de5701f3bac4bc520b750bb60694571ec7", "a76a874bada2291e9436d8c7738d46a7ca7039c0bb20a1741569983bcf33c842", "b6315d2a021e1515985121904a679fe5e5301e98a4f72faa66919563be5bba84" ] }, { "id": "W42-23", "period": "2026-10-15", "observation_zh": "小模型流程跑通以後，我再把入口接回原本 session 持有的模型，核對原參數物件和更新旗標沒有被換掉。更新、保存和兩個新程序的下一步有實際一起跑過，但仍是縮小 CPU 案例。", "receipt_sha256_commitments": [ "05b14faedae66bc92fbfae589a8affbad1be067a63b11e8fb1f6fe3fc809e773", "f4eff02665dffdc90c4f01dc664a64a711384faa436e96b6bc9dcf2bdee6be65", "f9bbad36d38ecfb3d3d92ee3bf565779f8ec78008fa39e9e961a1acbb4ff01c5" ] }, { "id": "W42-24", "period": "2026-10-16", "observation_zh": "10 月 16 日，我先在縮小的 CPU 模型上，把 Python 更新、Group、保存和換新程序接著算串起來；state-value 與 BF16 儲存的限定案例都有核對，但還不能把它當成完整 9B 的驗收。", "receipt_sha256_commitments": [ "3ccfad7ca082008025abde7487fa0c9c99ceb532c66ab3371dccfdb11adf2fb6", "7f7277bd1a1aef04d108c818ebb138f23d2d98bade1377fd527f635020b8239b", "81e593141e7fe43d991d66e513e4fcd58beb2a29221fc8a2b7a1397afe322b22", "e80919a3676aecc7e06692211444d27085e4115bff43811b6054ba4bda873de3" ] }, { "id": "W42-25", "period": "2026-10-16", "observation_zh": "這幾天也把部分梯度、回滾備份和雜湊改成分塊或寫到磁碟，讓成本比較能被檢查；不過模型、優化器狀態和保存時的部分完整物化仍然在，記憶體問題還沒有全部收完。", "receipt_sha256_commitments": [ "301891106c21c6d9bfa7b546a008f8b55e7d5d42c5e304dc216c7853a1b13bf7", "3ccfad7ca082008025abde7487fa0c9c99ceb532c66ab3371dccfdb11adf2fb6", "ba55d2d5a165adbce168ee37a96125b0bd3121db7cf5ce08db0ba0349860fcc6", "c1c9eb8c34d2b62c4e3436f1081f60706b46633674145eaa9a37cb03822cbbea" ] }, { "id": "W42-26", "period": "2026-10-16", "observation_zh": "我也另外檢查更新寫到一半才失敗會怎樣；在縮小 FP32 案例裡，確實讓第三次寫入之後出錯，再核對回滾和新程序接續，而不只是測還沒改到資料就拒絕。", "receipt_sha256_commitments": [ "24640dfc1b7f3692f1f7ae6fdde031f46c27e1e5f31c3e12aacc5a368f22303a", "49cd4f8f491ba584c5be0503d8a83944de3b40b0c6e4f3df3f1d3d8b64df7ac2", "cf2ca5583d9df25384defc3d83789e35803d6f62f90fc97fdd73fa1e14dd4a06", "e403b9f50e794aef416322b8964fe87f15947a03e1ae7b97f5f91b76d76caf78" ] }, { "id": "W42-27", "period": "2026-10-16", "observation_zh": "BF16 的部分也從空的研究情境往非空情境多驗了一段，不過混合精度是明講的計算規則；較寬配置目前只有局部數值檢查，沒有把它寫成整套冷啟動都通過。", "receipt_sha256_commitments": [ "23d641b6ecf5a3490dd23daaa83bc856bcf2701fdf0003ad242bca1c9de33b60", "24640dfc1b7f3692f1f7ae6fdde031f46c27e1e5f31c3e12aacc5a368f22303a", "2973623e3ded9a651b02b016bad7629a8405647740880b6077cc4905ae28dbc0", "e478ff151a9e69cf28d518c3a8bdcd30e2e6752e34fe52fb620468b58d41bc31" ] }, { "id": "W42-28", "period": "2026-10-16", "observation_zh": "保存點的部分，我也用指定的舊保存點換目錄、換新程序接著算，核對列出的狀態；這是在同一台電腦上做的搬移，還沒有變成跨機器或任意保存點都能恢復。", "receipt_sha256_commitments": [ "24640dfc1b7f3692f1f7ae6fdde031f46c27e1e5f31c3e12aacc5a368f22303a", "e4f0804a87b31a99a37b7a35f13c245dfd02ae8f2d0daaf46afb2cab0c9bf42f", "eb7f17e2a3818e148c73d7a0520717dc422610df34f1e8dab0e39763c5dc9947" ] }, { "id": "W42-29", "period": "2026-10-16", "observation_zh": "在上完整模型以前，我先把已知的權重和 buffer 成本列出來；本機 6 GiB 顯示卡連這一部分都放不下，未知成本也還在，所以這裡的結果是先拒絕執行，沒有硬把它跑成成功。", "receipt_sha256_commitments": [ "18834e40708aad314937ef7000efe39751db1d8f64514799acca62235b24055f", "24640dfc1b7f3692f1f7ae6fdde031f46c27e1e5f31c3e12aacc5a368f22303a", "cf2ca5583d9df25384defc3d83789e35803d6f62f90fc97fdd73fa1e14dd4a06" ] }, { "id": "W42-30", "period": "2026-10-17", "observation_zh": "另外做了一個離線的研究證據記憶工具，把原始收據、失敗和人工整理的筆記連在一起，讓之後查資料能回到原件；它驗的是資料保存與檢索，還不是小麥已經有了模型層的長期記憶。", "receipt_sha256_commitments": [ "0b38521a6f5c729341fc568373b0bc82644a2627cb4c69b713629ccfbfc1eebb", "555d19d060608ebe47c3f53669f662e37711049ee0673158b8e05f72e5409e69", "8da494649a8c993aa2d6b6aafd9d94e3233591912581d1f240fe3013d8d893aa", "e86cd032c01eb7709ee060de98102c530027fa8bbc90e0f9ee1c83b31f8ec20d" ] }, { "id": "W42-31", "period": "2026-10-17", "observation_zh": "10 月 17 日，我把入口、呼叫到的方法和保存恢復的範圍再拆開核對；有實際執行的部分就留下呼叫和新程序接續證據，原本沒有走到的生成迴圈和保存分支也沒有算進完成。", "receipt_sha256_commitments": [ "2c3e1298003ccf388af96c55aae9a52d488fdb9b15274aec7d39217979fa086f", "69611061be75b8d363600cd45fccfdad92fce1f756b546b76d8ea67bc966a2ae", "9afd0ac7dc97a7147c97a42e32be687bb3439eb90aac979b65e807119f0cebcd" ] }, { "id": "W42-32", "period": "2026-10-17", "observation_zh": "保存中途出錯也有另外驗，包含指標已經往前、但帳本還沒寫完的情況。這時預設載入會拒絕，需要明確指定上一個有效保存點；目前沒有把多檔案保存寫成已經具備原子性。", "receipt_sha256_commitments": [ "2c3e1298003ccf388af96c55aae9a52d488fdb9b15274aec7d39217979fa086f", "9e41ae7a31e5e7cc03bca79d0a8747d24cc372bbc770b03e849f0e23f8fecb18" ] }, { "id": "W42-33", "period": "2026-10-17", "observation_zh": "我也把指定的程式、依賴和保存資料放到另一個根目錄，用新程序繼續跑並對狀態；這次連 runtime 的位置也分開了，但仍然是同一台 Windows，沒有把它說成換機或重裝後都驗過。", "receipt_sha256_commitments": [ "2c3e1298003ccf388af96c55aae9a52d488fdb9b15274aec7d39217979fa086f", "d3f6775b2cb5d21870d44fa0020b84ead5f094c44176d00cd135f258a85ec6aa" ] }, { "id": "W42-34", "period": "2026-10-17", "observation_zh": "盤點時也開始把「程式有寫」、「入口接得到」、「真的呼叫過」和「數值核對過」分開，因為檔案放進交付包，還不能直接代表整條路徑已經跑過。", "receipt_sha256_commitments": [ "2c3e1298003ccf388af96c55aae9a52d488fdb9b15274aec7d39217979fa086f", "9afd0ac7dc97a7147c97a42e32be687bb3439eb90aac979b65e807119f0cebcd" ] }, { "id": "W42-35", "period": "2026-10-17–2026-10-18", "observation_zh": "正式入口的工作也往前接了一段：在 h8、三層的小模型與單張 GPU 上，更新、保存、換新程序恢復和再更新有了限定驗收。十項關卡裡有沿用不同來源版本的歷史證據，也有混合組合，並不是十項都在同一新版重新跑過。", "receipt_sha256_commitments": [ "0e5be6340f37a904b3777a4033b1e9982b89de5feb50823f48aca852ef2ad283", "17cfc791c9e272dda37e77d47ba532294bda414e09e1ba8b590d1ac967ceed4f", "ab590f357f6ec1a7d6dd668fd088eb750e7f581d4e30e40c7854af3a347eb616", "f4772e923939659d3eb23004ed088e81c1e998dd22951e7f5c217b11de57102c" ] }, { "id": "W42-36", "period": "2026-10-18", "observation_zh": "其中非空 state-value 的新版數值比較有通過，不過輸入也要照實寫：原文 491 個 token 只編入 486 個，末尾五個沒有進到這次範圍，所以還不能稱為完整輸入驗收。", "receipt_sha256_commitments": [ "5116fbbb277e300857dc4eeb93b50aeaaf895d8a7fe5fc4640e37c860df2ed91", "f4772e923939659d3eb23004ed088e81c1e998dd22951e7f5c217b11de57102c" ] }, { "id": "W42-37", "period": "2026-10-17–2026-10-18", "observation_zh": "接下來還要補的，也包括一般自由輸入的路徑，以及梯度和保存真正吃掉多少記憶體。現在 CUDA 梯度要先容得下選定的整批資料，再離開守門區寫到磁碟；保存也仍有完整張量和圖的物化，不能只因為有分塊，就說整體記憶體上限已經解決。", "receipt_sha256_commitments": [ "6194f4899b8c87f6ad97e6be4bf857134d45226e15d532402ba5d3684791b393", "8e9b6cf19183203bea5ec69a18728d04a2b38af67417aede178d96b2c7c8c238", "aa077c27b974564ef54fb985f0c5944a031e1674c887ca9163dfd70ab6467618", "eb9bcd3c5521a6a6f2d674b725cc72a1dfaaf63234dd32b3d639174b83116eba", "f4772e923939659d3eb23004ed088e81c1e998dd22951e7f5c217b11de57102c" ] }, { "id": "W42-38", "period": "2026-10-15", "observation_zh": "我也把先前分散的部件整理成身體拆解圖，方便回頭理解哪些是在縮小模型驗過、哪些還沒接到完整流程；圖本身是說明資料，沒有增加新的模型實驗結果。", "receipt_sha256_commitments": [ "4a499baaefb9343ab69a6992bd1595db24fdba81d724ac3e8929ed0674f3cfd2", "9a96854bb1faf0cf9d903d401e7125faf53e71493a0bb89205f4b2c37b79fe8f" ] }, { "id": "W42-39", "period": "2026-10-18", "observation_zh": "這週也先把原來保存的研究資料搬回新的運算環境，另外建立符合既有規則的執行環境，確認小型檢查能跑，再繼續接後面的模型流程。", "receipt_sha256_commitments": [ "89d3fcc6d9ab4a5cc11cd65c2d50705d69c5c74d91932a6bebbff84a1e32a154", "b4cc146f6e11e32abd977e124010322590f681e59fef64f120137fc485d6fd19", "e97985de530634b5d0aef20b0bd9c28352a156e2e3137cd0c72aa1a8c1c0965b" ] }, { "id": "W42-40", "period": "2026-10-18", "observation_zh": "在接到大模型以前，我先用小模型逐段比較新的運算路徑，也補上更新前後的輸入檢查；指定案例的梯度有對上，但這部分仍然是小模型的檢查。", "receipt_sha256_commitments": [ "55612a6a5e823410c315d5b220d0e75c902cc335bc256c52b083944719ff1c75", "63c60f3ba8a01d4cfcd701f7a8d829550a5e500a026d3f040218ca2d1415c7a7" ] }, { "id": "W42-41", "period": "2026-10-18", "observation_zh": "我也補了一個有記憶體上限的前綴暫存，讓重複計算可以少做一些；有空間才用，空間不夠就回到原來的路徑。開啟、關閉和回退這三種小模型案例，保存的梯度比較都有對上，實際速度還沒有另外量。", "receipt_sha256_commitments": [ "e0ce5d49e883401dcf11c72ace30c50f272a12cbfacf6856b321302ea587056c", "e3114adf39e5273b8768777e2b607171e40ed0e7c21b37c3895e8d9a41f6bb0f" ] }, { "id": "W42-42", "period": "2026-10-18", "observation_zh": "中間也有過外層程序正常結束，但模型入口其實失敗、沒有可驗收更新結果的情況，所以我把外層狀態和實際做成的事分開看，這些失敗也另外留著。", "receipt_sha256_commitments": [ "2a3e58b5dd273ac580a240a5c9cc1b82649ad04104e7945955f45e21fba44c28", "53ef1958a307fd8cea357c9c5fdb934f5015825d269ad8c0e37645821c91ae72" ] }, { "id": "W42-43", "period": "2026-10-18", "observation_zh": "A100上的9B等級模型，這次選定47組參數，第一次更新實際有6組參數改變；我把更新後的狀態存下來，接著拿去做新的程序恢復檢查。", "receipt_sha256_commitments": [ "a844fa7c78fafd213113baf07ef682d14dc7d92086b77d5ec6f3467eba952523", "afb0a4437128964c0d6c7b254b8e2b31977fed2687eda01a0073e19ac2bc6071" ] }, { "id": "W42-44", "period": "2026-10-18", "observation_zh": "換一個新的程序之後，載入的參數和優化器狀態，都能對回第一次更新後留下的投影，亂數狀態也另外核對過，這才讓我知道這一段可以接著往下做。", "receipt_sha256_commitments": [ "3bfd40198d6ba75e43e8a1f1b7efe6af62b6b3c535b97bc4f6d205a3173e3d8d", "9dca61e4eb5d9426b7a4937191b61552af8c591f57e9ceab14ab5644195719c5", "a844fa7c78fafd213113baf07ef682d14dc7d92086b77d5ec6f3467eba952523" ] }, { "id": "W42-45", "period": "2026-10-18", "observation_zh": "恢復後又做了第二次更新，這次有9組參數改變，也存成另一個保存點；不過第二次保存完以後，還沒有再換第三個程序載入，所以這一段仍然要接著驗。", "receipt_sha256_commitments": [ "9cfb8efbb4994e657c4945979117dd2321bef2f221f4c70e9f90446e9ef84768", "f99a36b3fafd8b9fcf30a8cada533dc3fea6162b2ecf5912b9ace4262216c95e" ] }, { "id": "W42-46", "period": "2026-10-18", "observation_zh": "原本第一輪的逾時和第一次恢復指標錯誤都還留在失敗紀錄裡，後來接續成功，也沒有回頭把那些結果改成通過。", "receipt_sha256_commitments": [ "59e9ad0074eeabbc2816751447355b1c62b48bca3990a318e35c14840fc12a41", "9cfb8efbb4994e657c4945979117dd2321bef2f221f4c70e9f90446e9ef84768" ] }, { "id": "W42-47", "period": "2026-10-18", "observation_zh": "這次文字仍然由主機固定指定，小麥沒有自己提出研究問題。更新和保存接起來了，但任意輸入、長輸入、數值等價和能力是否真的提升，仍然是接下來要分別驗的事。", "receipt_sha256_commitments": [ "9cfb8efbb4994e657c4945979117dd2321bef2f221f4c70e9f90446e9ef84768", "afb0a4437128964c0d6c7b254b8e2b31977fed2687eda01a0073e19ac2bc6071" ] }, { "id": "W42-48", "period": "2026-10-18", "observation_zh": "這週也把這輪程式、實驗紀錄和保存狀態整理到私有備份，做過封存內容的校驗。不過從完整備份重新解開、在另一個環境重啟，還沒有完整走過一次，所以我把資料保管和可重新執行分開記錄。", "receipt_sha256_commitments": [ "138ee0d2c62434e9581928906b3081cff960201f35d37c3f167df43cd91238de", "d21fa9ead69c4936eb27dfaf197912436138e709f3836ec6ed48a6f31ef8296f", "fdcb61ed2bf71a01157574b8896e5f1d8bf3e5713cf205d5f7c76bc3743d3b05" ] }, { "id": "W42-49", "period": "2026-10-13", "observation_zh": "10 月 13 日，我也把這輪已結束實驗的成功和失敗記錄一起封存，當時有核對封存與成員內容。這是在保住研究記錄，還不能拿來當模型能力通過。", "receipt_sha256_commitments": [ "1e4ffdbc329b89a904b6327026563b8082435bffd9db3aa592b7c71e86156233", "25b7d8a3851c4307f502678d435318c5c73dcf1f09e899a763d8ca549922ccdf", "2ea23b1cf022fd43cd566acc1fc43cacf2caf59da6c9d82b543d313a2dbe9ced", "75eab7a56882cbe1786fd50d3f4ce5ba90497e30ce9a0615b9de72411ef6a8b9", "7d6047f71decd561386646370bb8e5bfc5c22dd7fbffbccfd43b80350b2c0aa3", "dea805892c081091a994f313364bfb62f224f2629bc1b39bf2c234766780c673", "e6cd80cee35cd54c3f65137f974a6a04dda56f4fd065591342e172af49282556", "ec23bcc97cb5f276528cf01caaaf9ab0c9da73c8e7863e67189fe86f4d6dff9c", "f9b7f9536682db5374997e4985d46045dcd65ace9746fd79433eb42eb1142d62", "fc033587de9d373771814c3a6f2c67d349b4531b1b6cab5f41a7925009d45715" ] }, { "id": "W42-50", "period": "2026-10-15", "observation_zh": "公開之前，我先盤點已經放出去的內容，再把適合公開的結果和仍需保留的核心分開。", "receipt_sha256_commitments": [ "9eb1b7e24cbf26bc8bf2c5433b276c84ac49a1c55177e87661f3767591a851a9" ] }, { "id": "W42-51", "period": "2026-10-15", "observation_zh": "10 月 15 日，我把分散的研究整理成 38 個家族的公開目錄，也補上已複核的正面和負面結果，讓別人可以順著來源看。目錄收進來，和研究都通過，是兩件不同的事。", "receipt_sha256_commitments": [ "9c102347a7660e73363246443639778f709f7503f85ebc5ac4acad3d904a0159", "dfbef78f5a0cb72da29c63a31acfba1c02f9014f89cf08ab975c1947b430ec75", "e1a82570d78d944085d2f4accc56a5596586d774bd053e12bed8009fc4ec8bea" ] }, { "id": "W42-52", "period": "2026-10-15", "observation_zh": "我也把 GitHub 首頁重新整理成研究、證據、示範和文件幾個入口，讓第一次進來的人比較知道要從哪裡看。", "receipt_sha256_commitments": [ "a3221e5262822e9b409d8292564837d8059200535c1fc7fab919536e3bf59be8", "caadde5b87b7911a0030782d9e48fbad2e48c5996c423675e99a2edd5a7d0874", "e1198e64aa3c960bafe83d17645ea513f53e2dc59054456b67b07e2ef0d925c2" ] }, { "id": "W42-53", "period": "2026-10-17", "observation_zh": "這週也整理了一部分已核對相同的傳輸副本和舊影片副本，保留主份和原來的使用路徑。這是在減少重複保管，沒有增加模型能力。", "receipt_sha256_commitments": [ "4295f10180348e872cf54d2556e1695eb7a7aca7d61b798b855efe21a1207fe1", "85612710c08d4b851522adb5d2904d8b16f07d1c52504d5ad047b1b6735436f4" ] }, { "id": "W42-54", "period": "2026-10-17", "observation_zh": "10 月 17 日，我也整理了幾份歷史小麥封裝，留下完整基準和各版本真正不同的部分，並核對四個版本重建出來的內容。這裡核對的是檔案內容，重新啟動模型還是另外一項驗證。", "receipt_sha256_commitments": [ "8ccca1f312bda6db4861597864cb4fe720ef5485ab7efe0bf833c492b677fd7e", "bc744d2ac520b8724e33bb242881c9cba76cd94581a1ee575ee3244f48feab77", "c389f049d6edc20f0baacdcc0cd02105ebd6bb85febfa548fb4dc1bc29339797" ] }, { "id": "W42-55", "period": "2026-10-17", "observation_zh": "還有一部分時間花在把舊實驗留下來的問題、方法、結果和失敗整理清楚，再決定哪些舊版本可以退役。部分原始過程不再全留，這也代表那些歷史執行不能再逐值重驗，不能把一張結果卡當成完整恢復包。", "receipt_sha256_commitments": [ "4d1e1867e3fdd62144d2e168bcfd03254892f0f8f345f4edd4a3bf09b2ab1f36", "942f92a1a0f7360ad428dbbae7cb447789c31ced1f4e969024d1fa7be04d86be", "bc744d2ac520b8724e33bb242881c9cba76cd94581a1ee575ee3244f48feab77" ] }, { "id": "W42-56", "period": "2026-10-17", "observation_zh": "重複的共用依賴也整理成保留一份主資料、沿用原路徑的方式；這次沒有重新安裝環境來驗證完整執行。", "receipt_sha256_commitments": [ "353084e3d2aa4ca857438774fad336686391d0228ee550de3fc673ee5b544113", "4937d095559dfc7f4008fffac69604f93d31face4324e09f33a396af97b43bd0", "bc744d2ac520b8724e33bb242881c9cba76cd94581a1ee575ee3244f48feab77" ] }, { "id": "W42-57", "period": "2026-10-18", "observation_zh": "10 月 18 日，我把這輪可以公開的數字和限制補進公開倉庫，也實際讀回確認內容一致；核心、權重和私人保存點仍然保留。", "receipt_sha256_commitments": [ "134f6502098bc9bce877b385815e8f35b3e359282039ab2497967a16d01c52ef", "2714e2425c000e4eea17c4c7abf473c8b3c23c98e81de988da54d7816b519906" ] }, { "id": "W42-58", "period": "2026-10-18", "observation_zh": "網站這邊，我也把第 2 期週報接回原本的入口，中英文都一起更新，並保留前一版和這次增補的日期，讓後來補進的結果不會和原始記錄混在一起。", "receipt_sha256_commitments": [ "8c9e6f401ee118ed4f93afda080c81a06fd5d3e527d6acc701c4a8a9a57842d8", "afc7d15e2ea3ba7596ec905b2a86aea3bac331f1f445fd72026adbaa2bfe4f7b", "bc3bccd8d01a4aae61b100486c7c6d313853caa53a8f1d5210c2aa6dd341aa97" ] } ] }