v0.66.1 · 保留实验记录与运行前版本核验 / Evidence isolation and runtime preflight

重复运行抓梗实验时，每次都会保留独立记录；显式指定已有目录会提前报错，中断后的新来源清单不会再与旧成功报告混在一起。README 和中英说明提供首次运行、再次运行及独立检查命令。

开发提交会在长实验前核验官方稳定版本及 wheel 身份，结束时仍复核完整记录。新增显式 MuJoCo 3.17.0 候选配置，搭配独立安装的 SuperDex 1.0.1 FP64；历史安装默认值和实验记录保持原样。薄壳适配器记录实际版本，并拒绝 API/FP64 包版本不一致。

Repeated apple-stem launches now preserve independent records. Explicit nonempty outputs are rejected before native startup, including CLI and reset paths. Bilingual instructions cover first runs, reruns and independent verification.

Submission checks current official versions and wheel identities before expensive work, then rechecks raw records at completion. The explicit MuJoCo 3.17.0 candidate was prepared with independent SuperDex 1.0.1 FP64 packages. Historical installation defaults and records are preserved; the thin-shell adapter records its actual version and rejects mismatched packages.

Validation details and source identities are in the attached JSON receipt and SHA256SUMS. This is a scripted known-pose grasp regression, not hardware validation, visual control or a general solver-failure result. The asset contains summaries and identities, not an offline raw-data bundle. Installer ownership (#245), persistent dispatch admission (#244) and subsequent research tasks remain separate work.

Refs #243, #253. Scheduled development remains paused. No native engine wheels or robot assets are redistributed; no PyPI publication.

Validated: 1,223 tests (27 conditional skips), both complete 14 s episodes, independent acceptance (23 MuJoCo / 22 SuperDex checks), official wheel/source identity and strict bilingual documentation. Merged tree ed1467f5fb0252bbcb67e76919e65e2f405fcdea; commit a196c7fa3cf3754cd2d0dfeb7b827c9f0bed4786. PR #254.

Wuji contact evidence v1: point versus torsional contact

Wuji 双指夹球的不可变原始记录与离线复算包。v3 点接触的保持位移为 2.734 mm；另行冻结的 v4 合成抗扭接触为 1.025 mm，通过原 2 mm 门槛。六项 v4 对照分别检验保持/主动释放、逐指卸载、张指、禁用接触与名义零摩擦。仍有缓慢漂移，只对冻结的两秒窗口成立；不证明求解器失败、真实材料准确性或手内技能。

This immutable data bundle retains nine native records across four development cohorts, all original failures, the 15 unexecuted historical controls, exact generated scenes, frozen source snapshots, source/asset identities, independent scores and plots, offline contact/angular diagnostics, and continuous saved-state replay. The original v1 release-predicate mistake and later net-force cancellation loophole are preserved alongside their corrections and regression evidence. The latest independent reports check individual contact forces; all nine native task verdicts remain unchanged. No raw states or frozen task thresholds were changed.

See the Chinese entry or English entry for CPU-only rescore commands. evidence-bundle.json provides archive and per-file SHA-256 checks; SHA256SUMS covers the two downloadable payloads. Licenses and provenance are inside the archive. Engine binaries are not redistributed. This is a data release, not a new DexLab software version.

RSS wiping evidence v1: offline latch / no-latch rescore

RSS wiping: one raw latch/no-latch pair / 一组可离线复算的擦拭证据

This data release publishes the first controlled original/no-latch pair from RSS commit 23417de77edf59d614531a833334ed4e2206fc2c. Download and verify it using the CPU quickstart. The reviewed local scorer needs only Python, NumPy and Matplotlib; no simulator or GPU is required. Native physics reruns have separate GPU/version prerequisites.

本数据版本公开首个受控 latch／无 latch 配对，含完整材料点、接触、观测、参数、源码、录像和许可。按 CPU 快速开始从原始状态独立复算，不需要仿真引擎或 GPU。

• 159 files; archive 131,967,207 bytes; extracted 366,260,887 bytes.
• Archive SHA-256: ed78d647100ec0965330a131f827336fc7b388c72315d8d51a858c91e2013a0a.
• evidence-bundle.json records each file's origin, license, size, hash, role and attempt.
• attempt-index.json records all 12 execution entries plus preparation/postprocessing failures. Complete raw is included for this pair only. Other raw records and cross-cohort interpretation remain tracked in #247.
• Three derived bytecode caches are excluded because they can contain host paths; their sources and original hashes are retained.
• Both records pass integrity checks. Original/no-latch holding-patch displacement is 0.060631 / 449.981051 mm; cleaning is 95.473251% / 39.917695%. The no-latch task failure is retained, not corrected by rescoring.
• This is one historical fixed-geometry comparison of assistance present/removed. It is not a general solver-failure claim, new physics run, hardware validation, or complete RSS reproduction.

See the English report / 中文报告. Per-file upstream notices remain in the bundle; no simulation-engine binaries are redistributed. This immutable data tag does not replace the current code release.

v0.66.0 · 豆腐刀面推动 / Visible knife-driven separation

豆腐两刀切完后，由刀面接触把左块推开，动态图直接显示接触力、动作阶段和各块位移。默认左块移动 11.128 mm；中、右块移动分别为 0.038、0.015 mm。没有直接材料速度增量。

After two cuts, the knife re-enters the left kerf, pushes through native face contact, and withdraws. Both particle and surface replays show measured Fx and centroid displacement. Default and half-step pass; the no-push control is rejected as expected; the historical impulse protocol remains reproducible.

README and bilingual comparison pages distinguish the authors' original final velocity input from DexLab's knife-contact input. Synthetic CPU TLMPM, not calibrated tofu or a controlled three-method accuracy ranking. The original cut/contact solver is unchanged; the new input begins at 4 s.

The release includes four portable raw-record archives, one media/evidence archive, validation.json and SHA256SUMS. Validation records the full repository gate, 141-frame measurement checks, exact reviewed tree, local/public browser QA scope and deployed revision/media readback. Physics attempts and the pre-half-step resource-budget correction are retained.

中文展示 · English · Protocol and evidence

Publication transport audit: the first attempt stopped after creating the draft because immediate list readback omitted it (34.404 s). The second reached its frozen 600 s upload limit; readback confirmed all seven assets were already uploaded with matching sizes and SHA-256 digests. Publication resumed from that existing draft without reuploading or replacing assets. The validation manifest was frozen before upload; these delivery attempts did not rerun or alter physics evidence.

归档传输记录：首次草稿读取未及时返回，第二次上传触及既定 600 秒上限。随后核实七个文件均已完整接收且哈希一致，仅恢复发布步骤；没有重复上传、覆盖历史资产或重跑物理实验。

v0.65.0 · 豆腐交叉浅切与数值对照 / Cross scoring and measured comparison

README 现在直接展示三方数值对照，研究定位与目标位于开头。豆腐交叉浅切新增两条正交划痕、底部连通性验收、表面／粒子切换及带数值 GIF；原有连续两刀采用统一灰底与 16:10 构图。

The README now leads with the research mission and shows numerical values for BladeMaster, CRESSim-MPM and DexLab with explicit provenance. Cross scoring adds two orthogonal shallow incisions, intact-base checks and measured surface/particle replays. The original sequential-cut record is rerendered in a neutral 16:10 presentation.

Assets retain all four new native records, frozen source, independent scores and replay measurements. validation.json binds the full repository gate, source tree, browser checks and archive hashes. SHA256SUMS identifies every asset. The sequential replay reuses the v0.64.0 physical record; it is not a new benchmark run.

Synthetic materials and guided strokes are not real-food calibration or a complete reproduction of the author's knife-tip path. Author settings remain unmatched (#200), so the numerical table is not a controlled ranking.

Verification note: local bilingual browser layout/playback and mobile checks passed. Live merged revision, both README files, six pages and 16 media/measurement artifacts were independently read back. Additional post-deployment browser screenshots timed out; this is retained as an uncompleted visual check, not marked passed.

v0.64.1 — LIBERO restored-pose OSC initialization

LIBERO native state restoration now explicitly selects osc-restored-pose-v1. The pinned OSC cache and Cartesian goals are synchronized to the restored pose before the first action. Reset nullspace and gripper memory remain unchanged and recorded; unavailable source controller history is not reconstructed. Admission rejects an absent or unsupported contract, and independent scoring checks the declared initialization without relabelling historical records.

The native diagnostic evidence covers 57 complete on/off/repeat runs across 19 exposed development demonstrations and 14 previously qualified task paths. Four of five selected original failures became native successes, one remained failed, and all numerical negatives were retained. Some source-position deviations increased. These selected cases do not establish a full-task success rate or physical accuracy. Complete affected-task/all-demonstration regression, remaining source-history attribution, all 130 native tasks and visual-policy evaluation remain open in #198, #181, #188 and #182.

中文：LIBERO 新初始化协议在首动作前同步恢复姿态对应的 OSC 缓存与目标，保留并记录复位时的零空间及夹爪记忆。新增准入与独立评分检查，旧结果保持原协议语义。57 次完整对照覆盖 14 条已资格化路径、19 条已暴露开发示范；保留仍失败的案例、全部数值负结果和增大的源轨迹偏差。完整受影响任务、全部示范、130 项原生范围及视觉策略评测继续开放。

Validation: 1,122 unit tests (1,096 passed, 26 pinned-runtime skips), strict bilingual docs, complete MuJoCo/SuperDex 14-second regressions with 23/22 checks, independent raw-record and official-package identity checks, and unchanged source/assets/dependency versions. The recovery gate and terminal verification, including the retained interrupted recovery, used 1,306.33 of the original 1,800 service seconds. Previous failures remain recorded.

PR #208 merged commit 8af5d4b665728741e76e5c03ceb6d8aea60d172f with tree 6446a22e056e2bcf4a49aa66eda80cbf41b2fa18, exactly matching the validated combined tree. The PR documentation build passed. Native LIBERO remains on its qualified MuJoCo 2.3.7 / robosuite 1.4.0 path; the separate repository gate uses the qualified MuJoCo 3.16.0 profile. Attached reports contain selected evidence and validation identities, not third-party datasets or weights.

Main documentation build and deployment passed; the live source revision and both language pages were read back. This does not claim browser visual QA.

v0.64.0 · 切割数值动图与切面精修 / Measured cutting replay

README 和中英文网站现在直接展示带测量值的豆腐、香蕉动图；三列方法对比仍统一为豆腐两刀，香蕉单独展示。

• 动图显示实际仿真时间、切割阶段、切口宽度、内部体积变化、切后位移及材料对刀具反力。表面／粒子共用记录，播放器控制条移到画面外。
• 去除旧版最大 2.358 mm 的显示平滑位移，保留香蕉切面边缘；修正变形后锚点归属错误，支持当前三条直线切口的局部开口与未切断连接。
• 新冻结批次包括细采样豆腐、半步长、未切、浅切和细采样香蕉。豆腐 5,184 个原生粒子形成三块，左块相对移动 2.398 mm；香蕉 2,480 个原生粒子最终形成四段。评分器和物理源码未改动，历史与失败记录保留。
• 564 个表面帧与原生连通图一致；303 个导出帧的切口、质心位移和反力经独立计算核验，体积另作三重积检查。严格双语构建、桌面／窄屏播放检查及正常完整提交验收绑定到本次代码树。

豆腐三列对比 · 香蕉切片 · 论文细节与测量口径

Measured GIFs now lead the bilingual README/site. Values come from saved native states; original author clips remain externally hosted and attributed. Sharp cut rims, history-aware partial seams and component-aware anchor ownership replace the old smoothed banana display. Player controls no longer obscure telemetry.

The five-run cohort preserves unchanged scoring, negative controls, complete raw data and source snapshots. Six portable archives include those runs and measured media; SHA256SUMS and validation.json document the verified assets and exact-tree acceptance. v0.62.0 records remain unchanged.

GitHub visual readback exposed horizontal clipping in the README tables. A two-file width-only follow-up passed its own full unit suite and strict bilingual build, with executable code and media unchanged from the complete normal gate. Its initial wrapper-profile mismatch and corrected successful attempt remain recorded in validation.json.

This is a synthetic CPU implementation with independently defined estimators. The authored skin is an approximation for supported straight cuts, not the paper's linked-volume dual contouring. Banana plasticity/material and force qualification remain open in #204; missing matched-comparison settings remain in #200. No real-food calibration, full BladeMaster reproduction, same-condition method ranking or PyPI publication is claimed.

v0.63.0 — MuJoCo 3.16.0 native qualification

MuJoCo 3.16.0 qualification

Add an explicit 3.16.0 candidate profile and extend the existing discrete flex integration migration. Historical defaults and LIBERO's MuJoCo 2.3.7 path remain fixed. Preserve the pre-fix compiler failure, XML export rounding and complete passive-cloth trace with its first sampled midsurface intrusion; material and full cloth grasp remain unqualified.

Validated tree: 5cb88aaa24b998952d9a2fd0c2378844f93fe09c. Merged commit: 9f1836a567aa9f7cf252cf59c0ac8f2f7a753595. PR #205.

• Full suite: 1,105 checks, 1,081 passed and 24 pinned-runtime skips; strict Chinese/English documentation builds passed.
• Complete 14-second development regressions: MuJoCo 23/23 and SuperDex 22/22 independent checks; raw-record rescoring, official archive-to-installed-code identity and latest-stable admission passed.
• Cloth diagnostic: all 4,000 steps / 4,001 finite frames retained; first sampled midsurface intrusion at 22.5 ms, 0.049681 mm. The offline audit reproduced exactly. This is not a maximum depth or a continuous-time onset bound.
• Full gate and final verification: 1031.86 service seconds, peak 18,760,544,256 bytes, within the frozen budget and resource limits. Previous failures and their costs remain recorded. Setup, recording and auditing are included; these figures are not engine-speed comparisons.

Attached reports retain versions, source/record hashes, scoped checks, failures and limitations. They contain no robot/model asset redistribution. This is a GitHub code release; LIBERO initialization repair #198 and evaluation of all 130 tasks and all demonstrations continue separately.

中文

显式加入 MuJoCo 3.16.0 候选配置并沿用已有 discrete 弹性积分迁移，历史默认值与 LIBERO 2.3.7 路径不变。完整单元测试、严格双语构建、两个后端各14秒回归及原始记录／官方包准入通过。保留编译失败、XML舍入、完整静置轨迹及首次采样侵入；不把这些结果写成材料或完整抓取资格。附带小型核验报告与校验和，不重新分发模型资产；全部 130 项 LIBERO 及全部示范范围继续保留。

v0.62.0 · 豆腐三列对比与香蕉切片 / Recorded cutting showcase

README 与中英文首页现在先展示实际保存状态的切割回放。

• 三列对比统一为 BladeMaster / CRESSim-MPM / DexLab：豆腐连续两刀，撤刀后左块独立运动。前两列引用 BladeMaster 作者原站视频并署名；第三列是本项目运行结果。香蕉连续三刀单独展示。
• 新豆腐任务形成三等份，左块相对移动 2.799 mm，中、右块漂移均小于 0.08 mm；默认、半步长与细网格通过，未切/部分切负对照保持一块/两块。
• 香蕉保留三刀切成四段的物理验收，增加嵌入保存形变的平滑表面。表面为合成几何，非扫描资产；未完全切开的切口请看粒子视图。没有手工增加块体位移或缝隙。
• 表面/粒子切换、原站视频共同时轴、手机单列布局；README 首屏 GIF。

三列豆腐对比 · 香蕉实验

The three-column comparison now matches the two-cut tofu task in both author videos. Banana slicing is a separate experiment. External videos remain attributed and hosted by the original authors. DexLab's native recorded run preserves its own physical clock.

Persistent cut histories, blade/material reaction feedback, progressive component scoring, negative controls and original-task regressions accompany surface/particle replays. Full repository tests, strict bilingual documentation, both complete 14 s MuJoCo/SuperDex gates, independent acceptance, official-wheel qualification and browser checks are recorded in the validation asset.

Evidence assets retain all 25 cutting invocations, including development failures and source snapshots. Three archives split development (001–012), banana qualification (013–018), and tofu comparison plus regressions (019–025). Each file is verified byte-for-byte and every archived NumPy array is read offline. SHA-256 hashes cover archives, validation and replays. Historical v0.61.0 assets remain unchanged.

Synthetic parameters only: no measured food calibration, learned robotic policy, cutting-force convergence or same-condition method ranking is claimed. Complete author configurations and their gap/volume metrics have not been reproduced. Authored surface details are visualization, not physics evidence. No PyPI release.

Controlled-comparison inputs and metric gaps are tracked separately in #200; no new benchmark has been started.

v0.61.0 · 豆腐状软块对半切 / Synthetic tofu halving

豆腐状软块对半切 / Synthetic tofu halving

新增依据 BladeMaster v2 独立实现的单次切割 CPU TLMPM 参考任务：动态刀具与材料反力反馈、持久切面标签、撤刀后独立运动诊断，以及完整原始记录和独立评分器。

The default task creates two components with equal 50% mass. A one-sided diagnostic impulse after blade withdrawal produces 3.221 mm relative displacement. Both uncut and partial-cut controls remain connected and are rejected. Half-timestep and refined-grid cases also separate; the peak reaction is timestep-sensitive, so force convergence and real-tofu calibration are unproven. No robot arm, learned policy, GPU speed claim or full-paper reproduction is included.

• 中文实验说明
• English experiment
• Paper
• Implementation / validation: #195
• Merge commit / tree: baa4b48189ff52d9b7b8729ac25f538671707a9f / 8f17fb7d2bfbc97c4aa44788e9309011efd5e5b8
• Repository checks: 1,090 tests passed (21 existing conditional skips); strict bilingual docs; complete 14 s MuJoCo 3.15.0 and SuperDex 1.0.0 FP64 regressions with independent acceptance; official wheel/loaded-payload/raw/latest-stable qualification. The normal submit hit a TLS timeout in its last official-metadata lookup. The unchanged admission CLI completed that remaining check on identical-tree saved records; the failed invocation is retained. PR Documentation CI passed. See tofu-validation.json.

Attachments contain all eight cutting attempts, including the initial score-export failure and superseded development records; five final cases use frozen source hashes. SHA256SUMS identifies the raw archive and replay. The archive excludes third-party robot assets and private execution/control paths.

这是合成参数的首个切割实验，不代表真实豆腐材料或切割力标定。力值尚未建立收敛；四分之一切、多刀、表面重建和机械臂操作不在本版验收范围内。GitHub Release 不代表 PyPI 发布。