Paper readiness 与风险登记¶
当前判定:D3D-priority 文档终审、strict build 与 latest-build 公网 marker 验证 PASS;Formal experiment / private test No-Go。 现有数字属于 pilot/dev/historical, 不能填入 main paper tables。文档完成不等于实验完成;只有 28-day plan 的 data/metric/adapter/statistics/method gates 全过后才产生 formal numbers。
页面职责与写作来源¶
| Paper section | Canonical source | 当前缺口 |
|---|---|---|
| Problem/claims | Research story、实验总览 | 最终机制与 claim margin 人工签名 |
| Related Work | Related-work matrix、两套 metric review | citation-level source check 随 paper draft 再做 |
| Method | Method design | final ours mechanism/config |
| Experiments | Experiments、Evaluation protocol | frozen manifests/metric/budget |
| Tables | Paper tables | 所有数字保持空白直至 aggregator 填充 |
| Results | Results | L2 formal matched results、CI、coverage、failure denominator |
| Reproduction | Evidence、Project Guide | formal hashes、fresh reproduction、licenses |
Formal readiness matrix¶
| Gate | 当前状态 | 通过条件 | Owner / next action |
|---|---|---|---|
| Docs | PASS · latest build public | 本轮独立 review P0=0/P1=0;strict build、递归内部链接与 latest-build 公网逐页 marker 复验通过 | Paper/review track |
| Metric | No-Go | 人签 Physics outcome/wrapper/threshold/margin;contact-qualified secondary 只有 pairwise telemetry 通过后才能启用;Recon panel protocols冻结 | Human metric committee + evaluator |
| Data | No-Go | P0 D3D frozen manifest + Physics-reference manifest;ARCTIC RGB/VideoGen 只有激活时才需各自 Recon manifest | Data custodian |
| Recon baselines | No-Go | Table 1 external/proposed rows 为 D3D-HOI author-released native GT-CAD + Ours;Ours w/o HOI Align. 单列 paired ablation;3DADN 取得逐视频预测并统一重评后才进 main,否则 reported-only;CubeOPT reproduction 仅作 appendix control;ARCTIC rows 过 gate 后填数 | Baseline/integration |
| Physics baselines | No-Go | CoDA/InterMimic/RePHO 的 10-epoch pipeline validation 已通过;仍需作者预算 full train,以及各方法 clean reload、每 case 1 env × 1 canonical rollout/common eval | Baseline/integration |
| Statistics | No-Go | manifest-driven ITT、real CAD-cluster paired bootstrap、continuous coverage、rad/m、failure taxonomy regression | Evaluator/statistics |
| Method | No-Go | w/o Physics 与三个当前 Physics candidate rows 对应冻结 Method,且 safety non-inferior;corruption 仅由 robustness claim 触发 | Method/Ours |
| Private test | Sealed / No-Go | upstream 全过、hash/claim签名、无 leakage | Benchmark custodian |
| Paper number | No-Go | 每个数字回链 manifest/run/eval/aggregation;P0=0 | Paper + reproducibility |
Independent review ledger(Academic Research Suite)¶
下表按独立问题簇去重;同一问题跨文件的表现不重复计数。Docs P0/P1 只评价研究设计与公开文档是否
准确、完整、内部一致;尚未实现的 experiment gate 单列在下一节,不能因“已经写进风险表”而冒充实验完成。
| Review track / round | Date | Docs P0 | Docs P1 | Closure / verdict |
|---|---|---|---|---|
| Physics methodology + Devil's Advocate · initial | 2026-08-13 | 6 | 14 | outcome/wrapper、task–metric match、contact shortcut、valid-prefix、rad/m、来源与实现真值逐项修复 |
| Physics · first re-review | 2026-08-13 | 0 | 7 | 边界公式、来源事实、因果措辞、telemetry/reset、LaTeX 与测试状态继续收口 |
| Physics · final re-review | 2026-08-13 | 0 | 0 | PASS;formal Physics 仍有 5 implementation gates + 1 human metric-signature gate |
| Reconstruction methodology/baseline · initial | 2026-08-13 | 7 | 10 | output/GT、official formulas、baseline overlap、公平性、ITT 与 table design 修复;另记 P2=4 |
| Reconstruction · post-review new issues | 2026-08-13 | 2 | 0 | partial-output completion contract 与 ARCTIC validation leakage 均修复;full-body GT residual reopen 已在同一问题簇闭环 |
| Reconstruction · final re-review | 2026-08-13 | 0 | 0 | PASS;formal Recon/private-test gates 仍为 No-Go |
| Reconstruction D3D-priority / ARCTIC-optional · initial | 2026-08-13 | 1 | 3 | D3D automatic 真值/ITT、author-fidelity reducer、独立状态和 table↔manifest mapping 已修复 |
| Reconstruction D3D-priority / ARCTIC-optional · final | 2026-08-13 | 0 | 0 | PASS;T1-A1 required,T1-A2 conditional,ARCTIC/Generated non-blocking |
| EIC + Methodology + Devil's Advocate · initial | 2026-08-13 | 8 | 12 | claim↔table、provenance、完整 reference、contact/manual leakage、metric/formula 与 output contract 修复 |
| EIC · re-review new issues | 2026-08-13 | 2 | 4 | ARCTIC formal split、strict-build link、Day-8/17 freeze、provider status、historical markers、exact table columns 均修复 |
| EIC · previous final local review | 2026-08-13 | 0 | 0 | 旧快照 PASS;D3D-priority 修订后必须重新复审,不能沿用该结论 |
| D3D-priority / ARCTIC-optional EIC/Methodology/DA · initial | 2026-08-13 | 3 | 7 | compact 正文表、D3D/ARCTIC E2E 与 manifest 分离、D3D automatic baseline 真值、optional/supporting 角色、claim scope、paired comparator 与 reducer/mapping 逐项修复 |
| D3D-priority / ARCTIC-optional EIC/Methodology/DA · final | 2026-08-13 | 0 | 0 | PASS;strict build + 103 HTML / 8,639 internal refs / 0 broken;latest-build public marker verification PASS |
| Recon–Physics coupled-table redesign · initial pressure test | 2026-08-13 | 0 | 2 | 修复“固定 Physics=强行共用权重”与“不配对的跨数据集 reference 差异可直接相减”两处归因风险 |
| Recon–Physics coupled-table redesign · final | 2026-08-13 | 0 | 0 | 旧版 PASS;2026-08-14 已由新的 main-story-first 四表结构取代 |
| Main-story-first four-table redesign | 2026-08-14 | 0 | 0 | Recon → D3D end-to-end → captured-reference control → ablation;docs/LaTeX 同步 |
| D3D native + external-result provenance correction | 2026-08-14 | 0 | 0 | D3D-HOI 256 条确认为作者发布;CubeOPT 239 条确认为我们的复现;3DADN raw predictions 缺失,禁止把 reported State 冒充统一重评结果 |
| Latest full-site EIC + Methodology + Devil's Advocate · initial | 2026-08-14 | 5 | 8 | 修复唯一 provisional primary、valid-prefix coverage、Physics rad/m、baseline/ablation 分离、必要的 sensitivity/shortcut evidence schema;另记 P2=4 |
| Latest full-site EIC + Methodology + Devil's Advocate · final local | 2026-08-14 | 0 | 0 | PASS;P2=0;按人审取舍删除预设 parity/per-cluster 大表;strict build + 104 HTML / 15,888 internal refs / 0 broken;latest build public marker PASS |
| Overleaf Reconstruction experiment schema · initial | 2026-08-15 | 0 | 4 | 补齐 D3D native GT-CAD 边界、3DADN partial-output/coverage、独立 Recon ablation 与 ARCTIC optional common-protocol gate |
| Overleaf Physics experiment schema · initial | 2026-08-15 | 3 | 5 | 修正 provisional primary、ITT/coverage、rad/m、grounding target、contact gate 与 HDMI time gate;并行合并后新增 P1=1 同轮修复 |
| Overleaf EIC / Methodology / Devil's Advocate · initial | 2026-08-15 | 3 | 5 | 修正四表 claim mapping、baseline/ablation、D3D core/ARCTIC optional、独立 GT target 与论文 appendix 范围 |
| Overleaf experiment schema · independent final re-review | 2026-08-15 | 0 | 0 | PASS;仅同步 scientific tables/protocol/formulas/citations;空结果、正文 TODO 与 formal-number gates 仍未完成 |
最终合并后又做了三次短回归:Physics、Reconstruction 与 EIC 的当前论文实验 schema 均为
P0/P1: 0/0。论文只保留四张正文 scientific tables、一个紧凑 baseline-scope appendix 和必要公式;
manifest、adapter parity、readiness checklist 与 per-cluster 管理表没有进入论文。该快照连同真实引用已同步到
Overleaf a3ec942。
三名 reviewer 均独立使用 Academic Research Suite 的 source verification、methodology、domain review、 Devil's Advocate 或 EIC synthesis 流程。通过表示“当前文档没有已知未解决 P0/P1”,不表示保证录用, 也不解除 formal experiment No-Go。
随后进行的 Academic Research Suite writing-quality 精简只删除重复解释,并把 related work、status、metric formula 和执行协议收回各自的 canonical 页面。Recon–Physics coupled-table redesign 进一步重排了正文 evidence schema,因此单独做了 methodology/Devil's-Advocate pressure test;它没有改动 formal numbers,也没有启动实验。
Latest publication verification¶
2026-08-15 最新 target build 已由 target checkout 的共享 docs server 公开,并在唯一公网入口 https://die-amber-next-foundations.trycloudflare.com/ 逐页验证首页、Experiments、Physics/Recon metric review、paper tables、28-day plan、Paper Readiness、Results 与 Evidence;公开页已出现 ARCTIC 10-epoch pipeline-validation 状态与 checkpoint hash。Local strict build 共 104 个 HTML,递归检查 9,456 个内部引用、0 broken。该验证只证明公开内容与当前设计一致,不解除 formal No-Go。
论文侧只同步读者需要看到的 scientific schema:四张正文表、baseline scope、完整 Recon 对齐消融以及
efficiency/failure/Physics diagnostics。Manifest、readiness、adapter parity、per-cluster 和 review ledger 不进入
论文。该版本已同步到 Overleaf main commit a3ec942。
P0/P1 risk ledger¶
文档审查发现的实现缺口不能通过措辞“修复”,但可以成为明确 No-Go gate;这样不会把计划伪装成完成。
| ID | Priority | Risk | 当前处理 |
|---|---|---|---|
| R-PHY-CONTACT | P0 formal | marginal-force proxy 不能证明 hand×target-link pair | 文档已披露;pairwise telemetry + negative controls 是 contact-qualified secondary 前置;primary 不含 contact qualification |
| R-PHY-PHASE | P0 formal | extractor 只支持 outbound + optional return | benchmark scope 限定;未来 manifest 显式 phase schema |
| R-STAT-CLUSTER | P0 formal | CLI object_instance=case,不是 CAD cluster;continuous/failure aggregate 缺 |
文档降级为 missing;实现+end-to-end test 后才 Go |
| R-FAIL-ONTOLOGY | P1 formal | Physics failure-stage ontology 的回归测试与当前实现分支相反;指定小集合当前 36 pass / 2 fail | 不改未人签的 metric 公式;先冻结 reached_stage/failure_stage 本体与真值表,修复后两个分支回归全通过才 Go |
| R-RECON-ITT | P0 formal | evaluator 可能只扫描已有 result,漏失败 | scheduled manifest-driven discovery 是 formal gate |
| R-RECON-CONTACT | P0 formal | contact label 是输入,mask/gap 可是自洽量 | output contract 已纠正;accuracy/diagnostic 分层 |
| R-CONTACT-AUTO | P0 formal | 当前 HOI-align 读人工确认 contact,与 automatic video pipeline claim 冲突 | Formal main 冻结 automatic provider/raw VLM revision+hash,private test 零人工介入;manual variant 单列 diagnostic。不过则收缩为 provided-contact task 并给所有 row 同权限 |
| R-ARCTIC-ACTIVATION | P1 optional | ARCTIC 不是 D3D core;半套数据/baseline/bridge 会制造不可解释表格 | baseline 阶段结束前同时检查 view/data、formal split/evaluator、MANO↔SMPL-X bridge、ArcticNet-SF、HOPformer、Ours overlap;任一失败整块不进 formal queue |
| R-RECON-BRIDGE | P1 optional | Ours 输出 SMPL-X,ARCTIC official task/baselines 使用 MANO;joint/vertex metric 不天然同构 | 只在 ARCTIC 激活时阻塞该 panel;冻结 bridge 前对应单元 Unavailable,D3D core 不受影响 |
| R-RECON-FULLBODY | P1 optional | ARCTIC 有 SMPL-X full-body MoCap GT,但 official leaderboard 只覆盖 MANO hands+object | 只在另过 camera/world + SAM-Body4D lifecycle gate 后启用 optional full-body block;否则删 body block并不作 human-body accuracy claim |
| R-ARCTIC-SPLIT | P1 optional | official validation 会用于 integration,不能冒充 untouched formal main | ARCTIC 激活时所有方法走同一 test/eval server 或 untouched subject holdout;不可得则整块不进 formal main |
| R-SAM-BODY4D | P1 optional | target wrapper/default config 存在,但 submodules/sam-body4d/ dependency 缺失 |
ARCTIC full-body block planned optional;不阻塞 D3D core |
| R-BASE-PHY | P0 formal | learned Physics rows 无完整 target lifecycle | readiness 表精确降级;baseline-first 补齐 |
| R-BASE-RECON | P0 formal | D3D official 256-case aggregate 已按固定 9/8/17 reducer 复核并填入论文;Ours/Ours w/o 尚无完整 256-case formal 输出;3DADN 逐视频 predictions 缺失且 reported State 使用不同帧子集;CubeOPT 现有结果是我们的复现 | Ours 与 Ours w/o 完整后必须复用 d3dhoi-native-gt-cad-v1;限时取得/生成 3DADN predictions 并统一重评,否则降为 appendix reported-only;CubeOPT 只作带 provenance 的 appendix control;ARCTIC 不阻塞 core |
| R-EVIDENCE | P1 | target-local 与只读历史产物混淆 | Results/Evidence 加 provenance;formal row 必须 target/hashed import |
| R-MARGIN | P1 | practically meaningful difference/non-inferiority margin 未定 | 保留人决策,不伪造数值;private test 前签名 |
| R-DATA | P1 | D3D contact/denominator 未冻结;ARCTIC RGB、VideoGen reference availability 未定 | 协议阶段先冻结 D3D core;baseline 阶段结束前对 ARCTIC 作 activate/delete;VideoGen 无 GT 时只保留 downstream/feasibility |
| R-MULT | P1 | secondary/ablation 多重比较 | one primary;key secondary order;ablation Holm |
Result status vocabulary¶
所有公开 Results 使用且只使用:
| Label | 含义 |
|---|---|
| Formal | frozen manifest/contract、完整 ITT、统计/provenance gate 均过 |
| Pilot / L1 | matched 小规模或部分生命周期,只支持可行性/调试 |
| Diagnostic | sanity、proxy、upper bound、failure analysis,不参与排名 |
| Historical | 旧 checkout/旧 evaluator/旧 protocol,仅追溯 |
旧 toy 结果不删除;移出正式入口并标为 pilot/historical。低 success 只描述当前 pipeline 难度,既不否定 未来方法,也不支持 superiority。
Novelty 与 claim 边界¶
相关工作已经覆盖 video reference、q tracking、contact reward、residual action、rigid HOI physics、 articulated hand manipulation 和 generative HOI。禁止声称未经证据支持的 “first”。可竞争的主线是: automatic Recon 到 passive articulated full-body Physics 的 matched closed loop、reconstructed-reference execution、 contact/action shortcut audits 和错误归因;最终措辞必须由 formal baseline/ablation 结果决定。
完成标准¶
本 docs goal 完成只表示研究设计、来源、空表、审查、网站闭环已完成,仍可保留上述需人签名的 metric 选择和未来 formal experiment gates。论文结果完成则额外要求所有核心表有 formal numbers、P0=0、P1 修复/登记、fresh reproduction、artifact/license/anonymity/AI-use checks 通过。