2026-10-14 paper delivery scope and cost gate¶
Every item in the requested batch needs a disposition, but documentation is not implementation or verification. The budget is one A100/A30 or a Mac. The user did not separately request high-cost full training.
| Item | Delivery | Cost and scope | Acceptance |
|---|---|---|---|
| WeGR, 2610.16979 | Bounded real public-data training and matched controls for CoCo-VAE, DHPO, MSFT | No GPT-2 weight download; smaller causal decoder and MovieLens replace private search logs. No production gain claim | Implemented; A100 three seeds in about 16.7 seconds, negative results retained, unified CUDA CLI verified |
| UniSGR-V2, 2610.16132 | Documentation, online evidence, and deferral | 2.42 billion examples, 64 MoE experts, and 1024-event histories. Too costly to reproduce | Resource, original figure, and deferral complete; algorithm not implemented |
| Snap DPA relevance, 2610.14550 | Documentation, online evidence, and unavailable-resource record | Gemini/GPT teachers and 10K human-reviewed private labels; unavailable and uncertain cost. No fabricated substitutes | Resource, original figure, and deferral complete; algorithm not implemented |
| JEV transfer, 2610.14112 | Recomputed calibration, risk coverage, and cascade from fixed public predictions, plus commands | No paid APIs or Jev-output training. Select routing threshold only on validation | Recomputation verified; 4,800 test, macro accuracy 79.98% |
| JEV request audit, 2610.14631 | Request interventions, paired probabilities and flips, commands | Full seven-checkpoint rerun deferred; no actual capability result without author artifacts | Protocol and end-to-end tests pass; no model capability measurement |
| MemCon, 2607.13591v3 | Memory controller, observable execution and feedback, tests and docs | Lightweight UCB without gold to evaluated policy; not full six-benchmark reproduction | CPU three-seed operation diagnostic and persistence tests pass |
| EmbeddingGemma 2 | Pinned official checkpoint retrieval and bounded public evaluation | About 1.5 GB and 740M parameters; disable audio tower for about 440M image/text inference. 16 images and 80 captions; no pretraining | Real A100 checkpoint passed in about 1.51 seconds; initial batch failure retained |
| EV-JEPA, 2610.14324 | Documentation, Meta source, and deferral | 128 GPUs, 300 epochs, large video corpus. Too costly to reproduce | Resource, original figure, and deferral complete; algorithm not implemented |
| LoOPSD, 2610.18162 | Documentation and deferral | Path enumeration and up to ten rounds, 200 updates per round, four responses per question, 1024 tokens. Too costly to reproduce | Resource, original figure, and deferral complete; algorithm not implemented |
| RSIPO, 2610.15772 | Documentation, RSI link, and deferral | 3B/7B multi-environment long traces, skill replay, and iterative training. Too costly to reproduce | Resource, original figure, and deferral complete; algorithm not implemented |
| Netflix SID AR vs MTP, 2610.15864 | Source review only, not industrial reproduction | §5 explicitly lacks online A/B. Below industrial bar; 8B training also too costly | Exclusion recorded; reopen with new production evidence |
Verification rules¶
- Executable paths need algorithm behavior tests. Never populate actual results from paper numbers.
- New CUDA paths need real NVIDIA runs and sanitized receipts. Do not fabricate GPU evidence for CPU paths.
- Deferred items do not enter executable adapters/evolve or show unmeasured results on the public dashboard.
- One PR follows completion of scope, catalogs, minimal install, quality checks, and docs build.
- Review watermark stays 2026-10-12 for this delivery. Implementations do not substitute for outstanding new-window coverage.
Read and run¶
- WeGR: Real three-arm training, negative results, and unified command.
- MemCon: Persistent memory controller and CPU diagnostics.
- JEV regional transfer: Fixed public predictions, validation-only selection, and source hashes.
- JEV request audit: Gold-free request compilation and probability/action changes, not model reruns.
- EmbeddingGemma 2: Real image/text checkpoint, bounded download, four-dimensional retrieval evaluation.
GPU evidence: WeGR and EmbeddingGemma 2. Label and counter review: boundary check.
The other five were documented and deferred, not unreported implementations. Original-scale training requires a new user request and cost review.
2026-10-15 cost-disposition update¶
At user direction, high-cost no-reproduction cases are review DONE. They are not pending GPU TODOs, do not block closure or watermark movement, and are not retried automatically. Algorithms remain unimplemented with unchanged evidence limits. The October 12 watermark at delivery reflected incomplete new-window coverage, not untrained high-cost items.