2026-10-12: 16 implementations and fixed-window review closeout¶
The user accepted all previous P0/P1 candidates and two October 8 omissions in one MR. In parallel, close four tracks, two discovery channels, and Google/Meta source review before advancing October 7 to October 12. Implementing the short list cannot substitute for reviewing the complete candidate pool.
Discovery froze at 2026-10-12T12:55:18+08:00 for October 8–12, with preceding announcement overlap and identifier-month catch-up. Coverage refers to actual sources at that time, not the end of the day or the whole web.
| Priority | Paper/facility | arXiv | Scope | Actual status |
|---|---|---|---|---|
| P0 | Evo² | 2610.13025 | Agent/RSI | Implemented; A100 SFT and four real GRPO updates; eight GSM search diagnostics; circle failures retained |
| P0 | FORTE | 2610.13806 | Industrial recommendation | Implemented; three A100 seeds all retained baseline and did not select offline features |
| P0 | RSI-Anything | 2610.13912 | Agent/RSI | Implemented; A100 predictor updates and pair gate; zero circle gain |
| P0 | BranchOPD | 2610.13481 | Post-training | Implemented; A100 12 branches, eight PBR constraints, actual updates |
| P0 | CAST OPD | 2610.13257 | Post-training | Implemented; A100 cross-tokenizer alignment on 512 tokens; separate cast-opd namespace |
| P1 | Road MOPD | 2610.13521 | Post-training | Implemented; A100 conflicts/projections across 39 tensors; arithmetic subdomains, not original expert teachers |
| P1 | MSAR | 2610.13630 | Foundation model | Implemented; three A100 seeds actual byte-model training; unmatched parameters preclude formal gains |
| P1 | Orthros | 2610.13735 | Foundation model | Implemented; three A100 seeds training/KV equivalence; negative NLL retained |
| P1 | DualQuant | 2610.13767 | Foundation model | Implemented; A100 all linear layers of public 135M weights; software quantization, not native FP4 acceleration |
| P1 | BoundaryEvo | 2610.13285 | Post-training | Implemented; A100 one actual stage-two update under original gates; second round skipped without eligible teacher traces; prior failures retained |
| P1 | JANUS | 2610.13194 | Agent/RSI | Implemented; A100 real SDPA search, execution, and skill evolution; negative execution and old eager OOM retained |
| P1 | T2Mem | 2610.13348 | Agent | Implemented; A100 all 98 sessions for two official queries; external token F1 zero |
| P1 | RATED | 2610.13889 | Agent tool retrieval | Implemented; A100 dual-encoder training and feedback bandit; original API missing, execution failures retained |
| P1 facility | Jev Agent workflow evaluation | 2610.13092 | Jev/Agent | Implemented; three A100 seeds actual checkpoint generation, cost accounting, and sandbox; not JeV weight reproduction |
| Omission | WOVEN | 2610.12417 | Multimodal | Implemented; A100 original images and transition training; negative test retained |
| Omission | Unsupervised Multimodal Reward Modeling | 2610.12707 | Multimodal/reward | Implemented; A100 frozen VLM probe, verifier bootstrapping/retraining; no best-of-N gains measured |
ECC 2610.13066 and Google's user simulator 2610.13617 need source material and are outside the 16 implementation commitments, but must receive explicit defer/exclude/accept decisions. Record final decisions for other recalled candidates too.
Required checks¶
- Preserve defining models, objectives, policies, and actual paths. Fixed candidates or fixture scores are not capability evidence.
- Chinese paper pages include complete metadata, methods, equations, architecture, original figures and results, runnable entry, and local scope.
- Separate L0/L1/L2 from paper capability gains; retain negatives, truncation, failures, and no-op updates.
- Run CUDA paths on real A100/A30 with pinned public checkpoints/data and source commits, committing sanitized receipts only.
- Controllers cannot read gold and test cannot select. Evolve integration requires actual execution and reevaluation.
- Leave the dirty original checkout and
uv.lockuntouched, and deliver one reviewable MR without merge/deploy. - Move the watermark only after both channels across four tracks, candidate review, official Google/Meta sources, and nonblocking exceptions pass coverage audit.
Resume point¶
Branch codex/oct12-papers-and-review builds on origin/main including merged PR #199. Sixteen bounded implementations and actual GPU runs are complete, not sixteen full capability reproductions. Closed source review advanced the watermark to October 12. Other reviewed candidates remain independently tracked.
Label and counter audit¶
fidelity-source-inventory.json indexes review, not capability evidence.
| Path | Allowed use | Prohibited use and checks |
|---|---|---|
GSMHarnessEvaluator |
Official train-only validation answers in external numeric scorer; executor gets question and generated PROMPT | Evolution receives question/output SHA and correctness, not answers or raw executor text; test_oct12_agent_gsm_harness.py checks isolation and splits |
| T2Mem entry | memory.answer is generation; query answer is used after generation for external token F1 |
Do not pass labels into memory or generation; this is not paper LLM-judge accuracy |
| WOVEN/COCO loaders | WOVEN correct letter after rotation for training and post-output scoring; COCO captions for train/validation | woven_prompt receives question, choices, image only; split by public source group before training/layer selection; no test selection |
| BoundaryEvo | External analyzer can see train answers; fresh teacher outputs enter SFT only after external verified_reward checks |
Generators see no reference, no fabricated value on invalid analysis, and no stage two without passing; covered by test_oct12_post_training.py |
| Harness execution counters | calls counts attempts, alongside actual sandbox, frozen-weight digest, and output hashes |
Counts do not establish execution, gradients, or updates. Momentum drift with zero advantage is not GRPO learning. |
These are bounded L1 runs, not original capabilities automatically upgraded by seed count, receipts, or syntax audits.
Final local acceptance¶
python -m pytest -q: 1,857 passed, exit 0. Boundary was not excluded and assertions were not weakened.- Sixteen real GPU receipts validated; 93 execution-source SHA256 values match actual commits, not retrospectively filled from final HEAD.
compileall, Ruff, paper specs, fidelity inventory, minimal base-wheel install, and CLI manifest passed; registry has 387 adapters.- Schema-v2 metrics audited for 739 artifacts; platform completeness, strict paper coverage, and implementation reconciliation passed. This does not mean formal capability comparisons are complete.
- Manifest/catalog and 696 public dashboard entries regenerated without diff;
mkdocs build --strictpassed. - Caches, checkpoints, raw model outputs, machine connections, and
uv.lockexcluded. Local checks are not remote CI; use actual final PR checks.