Evidence：产物与数字如何复核¶
Evidence 不重新设计实验，也不把计划当结果。每一项先标 source checkout，再标证据等级；相对
output/... 只有在 target 中真实存在才称 target-local。
状态词¶
| Status | 含义 |
|---|---|
| Formal / L2+ | frozen manifest/metric、完整 ITT、statistics/provenance gate 通过 |
| Pilot / L1 | 小规模或 partial lifecycle，可验证接口/可行性 |
| Diagnostic / L0 | one-case/proxy/sanity，只定位问题 |
| Historical | 旧或只读 checkout、旧 protocol/evaluator |
| Missing | 当前 target 无可复核链 |
Target-local evidence map¶
| Topic | Status | Target path | 可支持 / 不可支持 |
|---|---|---|---|
| Ours Physics Full20 dev | Pilot | output/verification/shared_migration_20260801/full20_manipulation_3000_20260812_retry1/ |
部分 lifecycle/opening proxy/failure；非 formal task success |
| Ours train/reload/eval | Pilot | output/verification/shared_migration_20260801/ours_d3d_studio_residual_ppo_f/、ours_revalidation/ |
dev checkpoint/reload/eval；非 frozen split/budget |
| PHC-X common rollout | Diagnostic | output/verification/shared_migration_20260801/phcx_shared_tracking_smoke_d/eval/ |
human tracking interface sanity；非 learned manipulation |
| CoDA eval fragment | Diagnostic | output/verification/shared_migration_20260801/baseline_coda/eval/evaluation.json |
单一 dev eval；无 target full lifecycle provenance |
| CoDA author source provenance | Imported / source audit | submodules/PhysicalArtiObj-32102fe5a9933f87213fe480eb39a6fdddb918f5.zip、submodules/CoDA/studio_vendor_manifest.json |
官方 archive SHA-256 8de32a8b80d9df736fae178038b0b40dfaede43a3155bf7d76d452c74f68f5b6；vendor snapshot test 通过。只证明作者源与 Studio patch 可区分，不证明训练完成 |
| ARCTIC InterAct preprocess | Source/output audit | submodules/InterAct/data/arctic/sequences/s01/laptop_use_01/ |
当前仅 1 组，用于 pipeline validation；InterAct revision 96180a34f7b516e7f3520b853c19ea8679b8204f；human.npz SHA-256 93c3910e26b825dbfb51aac47b6c4b65f5e6c5d6a6d5ab71ec6464d2b121ae43，object.npz 6758baec9dad1683cc98c579fd16f5bd1e204f0c2d52552da74ae2c33e0ed129。ARCTIC-40 与完整 221 条 eligible case manifest 已冻结；其余 220 条的 InterAct outputs 尚未生成 |
| CoDA ARCTIC 10-epoch | Pipeline validation / L1 engineering | output/physics_only/arctic/coda/arctic_s02_microwave_use_02_pipeline_validation_10epoch/seed42/train/ |
256 env、10 epoch；checkpoint 可 torch.load，含 model+optimizer，epoch=10；SHA-256 93ba46fa8225e276b550a875c6c7eebc1978898a776ae55cf9cd97868c7a9bbe。作者 motion 输入 SHA-256 6541c90d37d59abc61442bd25034eb6b6725d042f007515933ff6d03967e50ff |
| CoDA epoch 500 reload | Budget-calibration evidence / diagnostic | output/physics_only/arctic/coda_eval/arctic_s02_microwave_use_02_checkpoint_500_eval/seed42/eval/ |
4096 env 训练 checkpoint 经作者 player clean reload，epoch=500、frame=65,536,000，SHA-256 f4c99222b2b2a74037777219f6ed5e7f20fd33d13d6fbaac07be52372f0727d1；common rollout 仅 14/300 帧并 early terminate，因此 500 epoch 明显不足，不能作为缩短预算建议 |
| CoDA epoch 2000 reload | Budget-calibration evidence / diagnostic | output/physics_only/arctic/coda_eval/arctic_s02_microwave_use_02_checkpoint_2000_eval/seed42/eval/ |
checkpoint epoch=2000、frame=262,144,000，SHA-256 1c5bbab67f645e5c9899ef405fb563adecf1eac24ae95660c887d7a1b17de820；作者 player clean reload，但 common rollout SHA-256 104d0d970ef6e17ccbf6975b20b97beaf25b6f9f6cb85a0cd4ef5b7cf0d99c7f，仍仅 14/300 帧，peak progress 与 epoch 500 相同。2000 未解决 early termination，须看 3000；在此之前 ARCTIC-40 bulk queue No-Go |
| InterMimic ARCTIC 10-epoch | Pipeline validation / L1 engineering | output/physics_only/arctic/intermimic/arctic_s01_laptop_use_01_pipeline_validation_10epoch/seed42/train/ |
256 env、从作者 sub2.pth 续训 10 epoch；checkpoint 可加载，epoch=13010；SHA-256 b243789f7e1b6a373d5847211bab6eecf9d80a26bcfcf7b7e265179d7ba8da8e。无 eval |
| InterMimic epoch 13250 reload | Budget-calibration evidence / diagnostic | output/physics_only/arctic/intermimic_eval/arctic_s01_laptop_use_01_checkpoint_13250_eval/seed42/eval/ |
4096 env 续训 checkpoint 经作者 player clean reload，epoch=13250、frame=32,243,712，SHA-256 b45182b9c01a2043747f2c21192705039f28d2e3474ad1498e24d8a59d034237；common rollout SHA-256 0d538b7b0ea584df003b493ea0006a195e62b1b54da41e99edbf0193998ed245，仅 13/653 帧并 early terminate。它证明 lifecycle 可用，但不能支持 250-epoch 缩短建议，仍须与 13500 比较 |
| InterMimic epoch 13500 reload | Budget-calibration evidence / diagnostic | output/physics_only/arctic/intermimic_eval/arctic_s01_laptop_use_01_checkpoint_13500_eval/seed42/eval/ |
checkpoint epoch=13500、frame=65,011,712，SHA-256 0555baddbbfa80ec306920c435f6e2539adba6fa8d9341f635240d968484ce43；作者 player clean reload，common rollout SHA-256 796d80fbe01923e5a808d406f405c403b5a95b40306b01d9a311770918845896。相较 13250，valid prefix 13→18/653 且 world MPJPE 322.8→135.7 mm，但仍 early terminate；500-epoch 续训不足以支持 full-horizon 或 ARCTIC-40 bulk launch |
| RePHO ARCTIC 10-epoch | Pipeline validation / L1 engineering | output/physics_only/arctic/repho/arctic_s01_laptop_use_01_pipeline_validation_10epoch/seed42/train/ |
128 env、双 worker 各 10 epoch；forward/backward checkpoint 均可加载，epoch=53010；SHA-256 21a9343cc4c5a3a0dac2a336565b7f475ca76b50a3f4b6183c657ef287383bf0 / 11e8586c56ae647f76a0e99e4b88fcff9e5568bb3fa56ae6fc824ad86e730d77。无 eval |
| RePHO epoch 53100 reload | Budget-calibration evidence / diagnostic | output/physics_only/arctic/repho_eval/arctic_s01_laptop_use_01_checkpoint_53100_eval/seed42/eval/ |
作者 forward/backward player 均 clean reload，common rollout/evaluation 写出；checkpoint SHA-256 f54ef1d2921553f18ece9725afb58e44efd93272734913b5490253ce2f3fe9dd / 0cd967fe577f341a274305ccd9d5b0a37f688891280973aa006ae3a5a1466fc1。仅执行 12/653 帧并 early terminate，因此只证明 lifecycle 可用，不支持缩短预算或性能主张 |
| RePHO formal artifact | Missing | 上述 pipeline-validation 路径 | 尚无作者预算 train → repair → clean reload → common eval，不能进入正式结果表 |
| HDMI | Missing | adapter/launcher/config code exists；submodules/HDMI dependency 与 lifecycle 缺失 |
不能进入 matched table |
| D3D-HOI official optimized results | Imported / not formal aggregate | baselines/d3d-hoi/；274 files、256 params.npy |
2026-08-14 已校验 official archive SHA-256 c5ae7e0b4d66c4f64b4c4757c5f5fd0ac0b52b7abb8b8b25f6229b6bd5f500b2，256 个 params.npy + 9 个 result.txt 逐文件一致；仍需 formal case manifest 与 evaluator aggregate |
| D3D v40 migration Full20 | Pilot/partial | output/runs/d3dhoi_shared_ground_v40_default_hoi_align_s100k_m10k_full20_v001/ |
20 result/case files，但 config/eval summary 不全；非 scheduled ITT/formal |
| D3D V41R Recon Full20 | Missing in target | V41R-specific manifest/aggregate only in read-only source | 旧 V41R 数字只能 historical |
| Formal Recon/Physics main tables | Missing | planned manifests listed in paper_tables.md |
数字格保持空白 |
ARCTIC full-env 训练时间校准（2026-08-15）¶
在 Tesla V100S 32GB 上保持作者环境数，只把训练截为 10 epoch。三个任务均完成并写出可加载 checkpoint。 实测 throughput 是可信的；完整训练轮数则必须区分“作者原值”和“本项目暂定值”。
| Method | 作者代码实际规定 | 10-epoch wall time | Steady time / epoch | 可支持的时间估算 | Observed GPU memory |
|---|---|---|---|---|---|
| CoDA | 4096 env，作者 mimicgen.yaml 固定 max_epochs: 3000 |
125.70 s | 7.785 s | 6.50 h/case，1 GPU | ~12.8 GB |
| InterMimic | 4096 env；作者 omomo.yaml 只有 max_epochs: 100000 上限并从 sub2.pth 恢复，没有规定 ARCTIC 固定续训轮数 |
338.68 s | 29.706 s | T(E) = (41.62 + 29.706E) / 3600 h；当前 E=500 仅是项目场景，约 4.14 h/case |
~23.1 GB |
| RePHO | 1024 env × 2 workers；作者从 epoch 53000 恢复，推理只选 epoch ≥54500，训练代码在 >56000 强制退出 |
250.89 s | 20.430 s（较慢的 forward） | 作者流程是约 1500–3000 个增量 epoch，即 8.52–17.04 h/case，2 GPUs；当前 E=2000 只是其中一个项目场景 |
~19.6 GB / GPU |
计时配置为 configs/experiments/physics_only/arctic/{coda,intermimic,repho}_full_env_timing_10epoch.yaml；
run records 位于对应 output/physics_only/arctic/<method>/*_full_env_timing_10epoch/seed42/train/。
因此，先前基于 3000/500/2000 全都属于“作者固定预算”而得到的 10,043 GPU-hours、15 张理论下限和
18 张建议值均已撤销。只有 CoDA 的 3000 可以直接外推；InterMimic 的 ARCTIC 续训预算必须先冻结，
RePHO 则应按作者的数据依赖停止/选点流程估区间。当前 target 还只有 1 组 CoDA 输入和 1 组 InterAct
human.npz + object.npz；ARCTIC-40 与 221 条 full eligible cohort 的 case manifest 已生成，但 bulk InterAct outputs
尚未生成。以上数据只能用于预算情景分析，不是正式结果。
Read-only reference evidence¶
以下路径可用于历史追溯，但不能暗示为 target-local：
| Topic | Read-only source | Status / requirement |
|---|---|---|
| D3D V41R reconstruction/tuning | /DATA/intern/hoi4d/projects/zhanbaiting/ArtHOI4D/output/runs/ |
Historical；需 source commit/diff + input/config/output/eval hashes 才能迁移为 evidence package |
| 2026-07-29 author-path rollouts | /DATA/intern/hoi4d/projects/zhanbaiting/ArtHOI4D/output/physics_studio_verification/final_author_20260729/ |
Historical interface witnesses；非 target formal rows |
| Source Setting-14 comparison | /mnt/cvda_mnt/xuyuan/project/intermimic-phcx/ |
Read-only method/source diagnostic；不是 target result |
Formal number lineage¶
每个 future paper number 必须能沿以下链复核：
claim/table cell
→ aggregation row + version/hash
→ scheduled manifest + split hash + CAD cluster
→ run_record + status/log + resolved config
→ checkpoint/common rollout/evaluation or result.pt/case metrics
→ input/asset/contact/code/evaluator hashes
缺任何节点则撤下数字或降为 pilot/historical；不得人工从日志拷一个看起来合理的数。
当前 formal blockers¶
- Physics metric 人工选择、pairwise contact telemetry、phase scope；
- Physics CLI 的真实 CAD cluster、continuous/coverage、rad/m 与 failure taxonomy aggregation；
- Recon scheduled-manifest discovery 和 panel-specific GT evaluator；
- D3D author-released 256 个结果已迁入 target；native evaluator aggregate、scheduled manifest 与 hash 仍是 core blocker。ARCTIC official baselines 与 Generated Video manifest 只在相应 optional/supporting panel 激活时阻塞该 panel，不阻塞 D3D core；
- learned Physics full-budget lifecycle / clean reload / 每 case 1 env × 1 canonical rollout；
- private-test seal 和 practically meaningful margin。
详细 gate 见 Paper Readiness，结果叙述见 Results，运行入口 见 Project Guide。