{"updatedAt":"2026-09-30","basis":"截至 2026-09-29 的本地交付摘要；網站製作時未重跑模型測試。","entries":[{"id":"cpu-regression","date":"2026-09-28","version":"v0.9.2","category":"verified","label":"限定驗證","title":"CPU 學習交易的回歸基礎","summary":"交付摘要記錄 4,555 項 CPU 回歸測試，失敗、錯誤與略過均為 0。","limit":"測試通過不代表自主研究能力已成立；信用 loss 仍是合成契約。","source":"v0.9.2 DELIVERY_STATUS.json / scope.full_regression"},{"id":"continuity","date":"2026-09-28","version":"v0.9.2","category":"verified","label":"限定驗證","title":"保存之後，下一步仍能對得上","summary":"限定 CPU 交易核對了主權重與動量保存，以及新程序接續的下一步一致性。","limit":"不是完整 True 整合模型的單檔冷啟動驗收。","source":"v0.9.2 DELIVERY_STATUS.json / new_process_next_step_equal"},{"id":"readonly-import","date":"2026-09-29","version":"True / A100 唯讀","category":"verified","label":"唯讀匯入驗證","title":"真實 True 的選定狀態匯入","summary":"A100 恢復同一個 True，唯讀核對 47 份主權重、20 份既有動量，涵蓋 1,067,974 個選定參數值。","limit":"此次沒有前向、loss、梯度更新或新 checkpoint；不能說新後端已完成真實模型學習。","source":"next_learning_v1 INDEPENDENT_ACCEPTANCE.json"},{"id":"real-layer","date":"2026-09-29","version":"v0.9.4","category":"implemented","label":"摘要記錄／限定比對","title":"真實第 0 層 GDN 核心比對","summary":"README 記錄 2 個 token 預填、1 個 token 接續解碼與 6 組梯度比較。","limit":"網站製作時未重跑 raw 證據；外層線性投影仍使用原生 PyTorch，完整模型整合待驗證。","source":"v0.9.4 README_zh.md / 驗收範圍"},{"id":"synthetic-loop","date":"2026-09-29","version":"v0.9.4","category":"implemented","label":"合成試驗","title":"串接修改、學習與恢復","summary":"README 記錄在 456 個合成參數試驗中，串接 Python 反向計算、AdamW、保存與新程序恢復。","limit":"真實 True 的新後端訓練尚未驗收；合成試驗不能代替真實模型驗證。","source":"v0.9.4 README_zh.md / 驗收範圍"},{"id":"core-goal","date":"2026-09-30","version":"核心研究","category":"pending","label":"尚未證明","title":"自己找到值得質疑的地方","summary":"自主形成與重新定義問題、提出新解釋，並保留可重現的研究能力提升，仍是待驗證的核心。","limit":"目前不能宣稱完整底座皆由 Python 運行、看清所有內部思考，或實現無上限自主進化。","source":"v0.9.2 autonomous_research_improvement_proven=false；v0.9.4 完整整合待驗證"}]}