原始完整判断

WeHub 本周真正变强在把 P02、TabU 与 TFM-Data 的职责边界压实，并把 TabU 从 Small-first 本地实现推进到可追溯的 full-table / savepoint / 数据供给合同；但这仍是已知表回顾评估与数据基建进步，不是未见世界泛化、P35 外部发布或 10 Paper full-paper gate 通过。

Evidence Boundary

  • Current fact：本期 canonical source 写入 /Users/cms/wehub-agent-mind/feedback-observations/2026-09-13-wehub-weekly-system-reflection.md，后续由 Team Dashboard system-reflection projection 渲染。
  • Current fact：本期读取了 WeHub FOUNDER_SEED.md、root README.md、hello_agent.md，system-reflection process.md，上一期 2026-09-06 report，Team OS state outputs，daily task-board projections，TabU / P02 / TFM-Data / 10 Paper / P35 source 与 receipts。
  • Current fact：方法版本仍为 v0.2；本报告继续区分 current fact、inference、risk 与 proposal。网页和 Discord 只作为 projection / feedback grow place，不自动升级稳定事实。
  • Current fact：/Users/cms/.openclaw/workspace 仍是大范围 dirty / mixed state；TabU、P02、10 Paper 的若干结论依赖局部 source、archive、receipt 和远端路径身份，不能被说成 clean release。
  • Current fact：root governance check 仍失败：缺 ripple/README.md，并存在 .DS_Store、.playwright-mcp、usl01-project-home-desktop.png 等 root ephemeral artifacts。
  • Inference：本周的主线不是单纯“TabU 分数更高”，而是研究系统开始把上游算子、表格模型、数据供给、评估口径和 checkpoint lineage 分开记录。这降低了 false promotion 风险。
  • Risk：如果把 retrospective same-table heldout、post-hoc +600 checkpoint、旧 Legacy TAR 高分或 OpenML12 已知表诊断写成泛化成功，系统会重新制造 capability inflation。

本周 Verified Deltas

1. P02 从旧 P02/RQ02 占位变成独立上游算子项目

  • Current fact：10 Paper paper-portfolio-v0.yaml 在 2026-09-09 记录 P02 slot replacement：P02 · Learnable Presence OTransformer，旧 P02-structural-mechanism-interface / RQ02 parked，新 workspace 为独立 Git repository。
  • Current fact：P02 hello_agent.md 明确目标是通用 OTransformer；TabU 是首个消费者之一，但 P02 不继承 TabU 或 P35 的代码、checkpoint 和既有实验。
  • Current fact：P02 已有 Gaussian 协议、native TabU-TAR 应用、positive-mass / headwise v2 边界、finite-guard savepoints、Medium / replay / full-table savepoints 等分层记录。
  • Conclusion：这是 10 Paper / TabU 生态的真实结构变化：P02 从“占位论文”转为可供 TabU 使用但不被 TabU 吞并的 foundational technology lane。

2. TabU 吸收 P02 反馈，但保持自己的模型与评估主线

  • Current fact：TabU README 已加入 P02 分工表，明确 TabU 负责表格基座模型，P02 负责 OTransformer，P02 native TabU-TAR 是独立应用分支。
  • Current fact：TabU 记录三条不能混同的实现/数学边界：TabU 数学采用各 head 共用 token-level presence；P02 native 采用 head-wise presence；旧 tabu.tar 保留原有算子。
  • Current fact：TabU 的 2026-09-13 P02 回流报告记录 full-table 协议：每张表全部 train rows 为 labeled context，全部 reserved-test rows 一起作为 query，truth 在模型外；12 张表共 7,894 query rows。
  • Conclusion：TabU 变强在 evaluation contract 和 lineage hygiene，而不是已经证明新 native 架构优于旧 TAR 或传统 baseline。

3. Full-table reserved 纠偏产生新的可比较面板

  • Current fact：P02 full-table reserved contract 纠正了旧 windowed / reduced-context 口径：完整表格推断使用全部原始 train rows 作为 context，全部 test rows 一次联合 query。
  • Current fact：owner-selected retained set 包含 Legacy TAR Small、Native Small identity/full/continuation、Native Medium replay3 start/+600、MLP、XGBoost。12 表宏平均 R²：Legacy TAR 0.78268253，Native Medium +600 0.77906846，MLP 0.77738705，XGBoost 0.78794406。
  • Current fact：+600 是终局后 owner 请求查看的适配性中间点；+1,200 完成但低于 +600。Legacy TAR 261760 是不同架构和历史训练线，2026-09-13 的 full-table score 是新评估，权重未变。
  • Boundary：这些是 retrospective same-table heldout results；reserved set 已参与 adaptation / selection，不能写成 blind validation、unseen-table ICL 或 architecture superiority。
  • Conclusion：这是本周最有价值的 representative work：系统把“能看全表才是合理 inference”转成了可复查的 evaluation contract，并保留更强的旧 endpoint，而不是强行拥抱新 native 叙事。

4. TFM-Data 被归位为平级数据供给项目

  • Current fact：TFM-Data README 在 2026-09-12 明确三方核心任务：P02 = OTransformer，TabU = 表格基座模型，TFM-Data = 支撑预训练和可信评测的数据供给。
  • Current fact：共同数据约定区分 source/DGP、world、typed table、episode、corpus/evaluation task，并明确 world split、表内行划分和 episode support/query 是不同层。
  • Current fact：生成器能力审计显示 Mixed SCM 可重建具体 world 并抽新行；但按行寻址重放、新持续供给入口、消费者 adapter 和训练/评估接入仍未实现。
  • Current fact：经历过 OpenML12 已知表训练的权重，在 OpenML-8 上只能作为 known-table diagnostic，不能再作为未见表确认集。
  • Conclusion：这是 system-level progress：数据供给终于从“历史固定表/Anchor120”转向 world-level supply and evaluation semantics，但仍处于 local_unissued / design-to-implementation transition。

5. 10 Paper / P11 / TFM public projection 有进展，但 core paper gate 仍失败

  • Current fact：10 Paper repo 本周出现 P02 replacement、P11 Tabular Foundation Models tree、TFM overview/map public deployment receipts，以及 P35 manuscript/release candidate updates。
  • Current fact：python3 scripts/check_portfolio.py 仍失败：旧 P02 structural-mechanism boundary file missing，USL-01 generated entry freshness marker drift，USL-01 ledger current_contract escapes project root，USL-01 generated page differs from dry render。
  • Current fact：P11 / TFM overview public receipts on 2026-09-12 are public projection and research-map improvements; they do not by themselves create a full-paper gate exit.
  • Conclusion：10 Paper 本周更像“portfolio identity and projection repair under active migration”，不是 full-paper production success。

6. P35 保持 release-candidate boundary

  • Current fact：P35 PUBLIC_RELEASE.md records p35-oattention-release-20260821-v1 as strict-verified candidate with 157 tests passed and an arXiv bundle compile, but manifest remains release_status: candidate.
  • Current fact：external release gate still requires clean reviewed Git commit/push, final repo/tag/DOI metadata, owner arXiv/license/author confirmation, rebuild from clean commit, strict verifier, extracted tests and isolated arXiv compilation.
  • Conclusion：P35 的 stage-closed evidence 保持有价值，但本周没有跨过 external release。

7. Team OS / Hosts / Providers

  • Current fact：2026-09-13 provider snapshot records hosts 8/9, monitored profiles 33, provider entries 312, primary usable / warn-or-unknown / down = 32 / 1 / 0, action items 9 with P0 0 / P1 1 / P2 8, routing-policy items 14.
  • Current fact：Agent TUI snapshot records 33/33 reachable/configured, attention 0, action items 0.
  • Current fact：agent-operability-conflicts-latest.md is stale from 2026-06-04 and should not override current provider/TUI snapshots.
  • Conclusion：Team OS support lane is usable and not currently blocking the reflection workflow; provider routing policy debt remains maintenance, not main battle.

8. Discord / AID / Human Attention

  • Current fact：Daily boards from 2026-09-07, 09-08, 09-11, 09-12 and 09-13 record P0 movement from TabU scratch contract → TFM contribution/evidence outline → TabU training migration/version compatibility → training receipt verification.
  • Current fact：AID daily partitions remain partial: some days read 0 usable Gong messages, 2026-09-10 has 11 usable messages, 2026-09-12 has 4 usable messages. These are bot-visible active-thread partitions, not full Discord/DM evidence.
  • Inference：Gong attention entered non-substitutable judgment at least twice: deciding P02 replacement / independent project identity, and correcting full-table inference / training migration framing.
  • Conclusion：Human attention was not consumed by generic system maintenance; it sharpened research/evaluation boundaries. This is a positive attention-quality signal.

Five Required Tests

Test 1 · Mainline and WIP Defense

  • Result：pass with boundary。
  • Evidence：P02 / TabU / TFM-Data roles are now explicitly separated in source; TabU README points to full-table evaluation and P02 feedback; daily boards keep training receipt verification as P0 rather than broad model expansion.
  • Boundary：10 Paper portfolio checks fail; root governance still fails; P35 not externally released.
  • Conclusion：主战役应是 TabU + P02 + TFM-Data 的 next credible evidence loop：from known-table receipt hygiene to independent world/data supply, not more same-table score polishing.

Test 2 · Evolution

  • Result：partial pass。
  • Evidence：上周 owner correction “不要补丁式修，要融入生长框架”在本周变成三层继承：P02 独立上游、TabU 保留模型主权、TFM-Data 平级供给。
  • Evidence：full-table correction reused prior checkpoints and legacy endpoint without overwriting them; old high score was retained rather than discarded for a newer narrative.
  • Boundary：stable second-use better outcome 仍只部分成立。新的 contracts 已被复用到 P02→TabU 回流，但还没有一个独立 agent 在 TFM-Data 新世界供给下完成更低成本、更好结果的 full evaluation。

Test 3 · Human Attention

  • Result：pass。
  • Evidence：Gong 的注意力进入了项目身份、评估口径和训练迁移判断，而不是进入 dashboard 或 provider 修修补补。
  • Risk：AID partial silence 仍可能制造 false negative；报告只把它作为 coverage boundary。

Test 4 · Complexity

  • Result：mixed。
  • Positive：把 P02 / TabU / TFM-Data 分开减少了 source ownership 混乱；full-table contract 减少了 evaluation apples-to-oranges；savepoint registry 保留 Legacy 和 Native 的不同身份。
  • Negative：10 Paper projection/ledger check failure、root governance failure、large dirty workspace、P02/TabU/P11 多条同名模型线会继续制造 complexity debt。
  • Conclusion：新增复杂度大多在解除真实瓶颈，但需要通过 check repair 和更窄的 next experiment 防止再次膨胀。

Test 5 · Representative Work

  • Result：pass with strict boundary。
  • Original proposition：TabU / P02 / TFM-Data together now express a more causally honest evidence pipeline: operator → model → data supply → evaluation identity。
  • Real Agent Society：multiple project surfaces exchanged roles without central platform rebuild; daily board, P02 source, TabU source, TFM-Data source and Team Dashboard projections stayed linked.
  • Flagship result：full-table reserved comparison plus savepoint preservation is a flagship evidence hygiene result, not a flagship generalization result.
  • External migration：not passed. Public TFM overview/map deployed, but no external scientific release/submission/generic user adoption proof.

主战役 / WIP 判断

  • Main battle：将 TabU 的下一证据循环从 same-table retrospective comparison 推到 TFM-Data-backed world-level supply and evaluation：先做 adapter / truth isolation / world split receipt，再讨论训练预算。
  • Keep defended：P02 as upstream OTransformer lane; TabU as model owner; TFM-Data as data/evaluation supply owner; P35 as release candidate; 10 Paper as portfolio/projector under repair.
  • Do not promote：Legacy TAR higher full-table score、Native +600、XGBoost macro best、OpenML8 diagnostics、TFM overview public page，均不能被写成 foundation generalization。
  • Postpone strategically：large pretraining, new public claims, P35 external submission, 10 Paper full-paper expansion, central metabolism/platform/dashboard rebuild。

Inheritance And Second Use

  • What improved：上周的 TAR / Small-first evidence lane 被本周继承成更严格的 full-table contract、savepoint retention 和 data-supply boundary。P02 的独立算子进展没有吞掉 TabU，而是成为 TabU 的 explicit upstream feedback source。
  • Second-use candidate：full-table planner / truth isolation / plan digest / immutable model checks can become reusable evaluation machinery for TabU and P02 native TAR.
  • Missing proof：还没有通过 TFM-Data 新世界供给完成一条 independent world / held-out world evaluation，并产生比旧 same-table workflow 更可信的 result。

Discord/AID 与 Human Attention

  • Current fact：本期 Discord writeback should be one concise summary after public verification, with one Gong question only.
  • Current fact：Daily boards exposed useful owner corrections, but AID coverage remains partial and cannot prove absence of attention on 0-message days.
  • Proposal：下周继续把 owner evidence 分为 Discord active-thread evidence、project source evidence 和 remote artifact receipt；不要合并成一个 attention counter。

复杂度与战略性推迟

  • Pay down now：10 Paper P02 replacement / USL-01 projection check failure; root governance hygiene if scoped separately; TFM-Data consumer adapter boundary.
  • Defer：do not build a central evidence database; do not start broad training just because savepoints exist; do not turn TFM overview public pages into a new dashboard program.
  • Preserve：P02 / TabU / TFM-Data separation is productive complexity. Collapsing them would reduce immediate friction but damage evidence integrity.

Directions 1-3

  1. Make next evidence loop world-level, not score-level：TFM-Data should hand one fixed-shape new-world supply slice to TabU/P02 with explicit world split, adapter, truth isolation and consumer receipt before any training claim.
  2. Repair portfolio projection as maintenance：fix the P02 replacement / USL-01 check drift in 10 Paper without reopening broad paper planning.
  3. Keep P35 and public pages bounded：P35 remains candidate until clean release gate passes; TFM/P11 public maps are useful routing surfaces, not research result claims.

Stable Mind Delta Candidates

  1. P02 / TabU / TFM-Data are three roles：P02 owns OTransformer, TabU owns the table foundation model, TFM-Data owns data supply and trusted evaluation; no result should silently cross these ownership lines.
  2. Full-table known-table evaluation is evidence hygiene, not generalization：full train rows as context and reserved rows as query corrects the task contract, but retrospective same-table scores cannot become unseen-world capability.
  3. Legacy endpoints deserve preservation：a stronger old checkpoint must be kept as a comparison object, not overwritten or dismissed because the current architecture is newer.

最值得 Gong 判断的问题

下周是否把主战役明确切到 TFM-Data → TabU/P02 的 first world-level supply + consumer receipt，并把 same-table checkpoint comparison、P35 release、10 Paper projection repair都降为边界维护？

本期状态摘要

  • Source：created at /Users/cms/wehub-agent-mind/feedback-observations/2026-09-13-wehub-weekly-system-reflection.md。
  • Public projection：pending render / deploy / public verification at source-write time。
  • Discord writeback：pending public verification at source-write time。
  • Final boundary：本报告是 feedback observation / proposal only；stable mind delta 仍需 Gong 判断后回填 canonical source。