October 23, 2026: Recent announcement and Google-priority backfill¶
This fixed batch includes eight low-cost components and one high-cost reference. Components passed frozen-source three-seed acceptance: OpenAttention, MoEZip, AOPD, GIFT, and PAIR ran on a real A100; Gemma protocol, Jev-DLC, and finite-policy RKL ran on CPU. Status is bounded_scope_complete. The global watermark remains 2026-10-15; this is not announcement-wide or window-wide full-text closure.
Frozen source: f3fbc6db4db084a2a6a4ee754f9a3d9e5b606ac7. Eight CPU diagnostics took about 9.56 seconds and five CUDA diagnostics about 1.73 seconds, excluding connection and sync, not full training times. Raw CPU, raw CUDA, and proof bind metrics and five sanitized receipts. scripts/build_oct23_priority_artifacts.py --check verifies raw hashes, exact seeds, fixed checks/metrics, catalog boundaries, and receipts offline; labels cannot upgrade evidence.
Fixed scope¶
| Paper | Bounded work | Classification and limits |
|---|---|---|
| Google Gemma four-arm protocol | Four arms, common holdout, BCa/TOST, and fixed author-log recomputation | September 25 backfill; CPU protocol, not new inference |
| OpenAttention | Single-head tensor mechanism | Cross-list; not full LLM or CUDA kernel performance |
| MoEZip | Routing-aware KV selection | Cross-list; fixed routing does not stand in for the full model |
| AOPD | Original advantage and distillation loss | Cross-list; no full rollout or LLM training |
| GIFT | Gradient-informed distillation filtering | Paper-specific method and resources; not full reasoning RL |
| PAIR | Intent and preference scoring interface | Industrial offline/online evidence separate from local mechanisms |
| Jev-DLC | Typed feedback decision updates | CPU mechanism, no large Jev checkpoint |
| Google collaboration RKL | Complete Algorithm 1 for finite MDP | Google coauthor priority; not ordinary token rKL or LLM training |
| OneLatent | Original methods, online evidence, and cost references | Full experiments too costly, not reproduced (review DONE), no retry |
The fixed scope and paper blocks record exact v1 times, primary first-author affiliations, code/license/revision, and stable keys. New URLs, new/cross-list entries, true first publication, backfill, and revisions are distinct. Gemma reruns and RKL LLM experiments retain their specific full_experiment_status rather than being forced into the high-cost category.
Sources and limits¶
Source receipt: eight October 23 arXiv announcements, 1,260 category entries and 916 identities; 499 new, 287 cross-list, and 474 replacement entries. These are not 916 new or fully reviewed papers. Exact v1 times were checked individually for selected work.
Google returned 29 pages/428 records and five new URLs versus the last successful 423-record capture; the failed intervening capture was not an empty catalog. Meta six pages/144 records and DeepMind nine/265 had no bounded URL difference, not zero organization-wide publications. Failed captures must report unknown/null.
| New Google catalog entry | Disposition |
|---|---|
| Words Speak Louder Than Order | Full text confirmed sole Google author; September 25 backfill included as protocol |
| Stakes-Adaptive Verification and Origin-Traced Rollback | Official TDC12058 public October 8; full text returned 403/HTML, not a PDF. Non-blocking archive exception without claims of method review or invented implementation |
| Regional Policies for Memory Tiering Systems | REGENT virtual-address-region memory tiers are outside this LLM/recommendation/Agent scope; no fabricated first-publication time |
| π-Flux Tuning of Quantum Paramagnetism... | Condensed-matter quantum physics, out of scope |
| Nature of the bosonic edge phase... | Condensed-matter quantum physics, out of scope |
RKL's Google collaboration was found through an independent full-text author check, without requiring a catalog entry or abstract affiliation. Targeted checks do not exhaust all coauthors. Do not recount ALPS, ASTEROID, CORD, CSD, Jev multi-correct options, AW-0, Right On Q, or FineSID, or count MemoProAgent v2 as a new paper.
Four other boundaries: Adobe LARA (2610.30423) reports observational production stratification, not confirmed controlled quantified A/B or approved full-traffic comparison. baseVultra (2610.30154) announcement/ID conflicts with the linked v1 August 23 date; do not force it into the day count. DConstrain (2610.30920) author information was unavailable and Verbalizing Activation Differences (2610.30934) lacks explicit affiliations; Google/Meta links are unknown, not disproven. None is automatically queued.
Four targeted post-training checks stay outside execution scope. MCPD (2610.30880) reports 4×A100 80GB and 1h36m for 100 steps; too costly, not reproduced, review DONE. Route-and-Generate (2610.30351) leaves the mean undefined at n_f=0 without located source. IRAH (2610.30167) cannot omit LLM reflective mutation and meta-learning warmup while preserving its core. HERD (2610.31401) reports 25 stages and 10,723 seconds; the paper's arr branch linked by the official README was not obtained and main is not a substitute. The last three do not claim completed review or expensive execution and are not automatically implemented or retried. See JSON for dates, resources, code, and unknowns.
Separate generation and acceptance¶
python scripts/build_oct23_source_audit.py /tmp/oct23-source-delta
python -m pytest tests/test_oct23_source_audit.py tests/test_gemma_priority_protocol.py
The source builder reads actual receipts, prior coverage, and the independent catalog to produce pending scope and append a discovery batch. It rejects an existing batch to preserve later acceptance. It does not go online, execute models, write runtime/metrics, or move watermarks. A separate curator finalization recorded frozen-source/GPU acceptance, not model capability from CPU multiseed checks.