Review through October 12, 2026: coverage, backfill, and exceptions¶
The central status review watermark advanced to 2026-10-12. Four tracks, two channels, per-ID dispositions, and official sources and exceptions passed audit_paper_coverage.py --strict. This closes a defined review window, not the entire web or all implementations. Search began at 12:55 Shanghai time on October 12, 2026 and covered October 8–12 plus announcement overlap and month-number backfill. It does not claim late-day, unpublished, or unindexed October 12 material. The next scan overlaps from October 11.
Counts¶
| Item | Count | Meaning |
|---|---|---|
| Deduplicated IDs across eight query groups | 1,597 | Four research tracks, each with submission-window and late-index channels |
| Inherited reviewed records | 1,019 | Not newly published items or new full-text reads |
| Newly reviewed query records | 578 | Each has an actual disposition; unavailable papers are deferred, not accepted as full-text reviews |
| Directed discoveries outside queries | 10 | Nine papers from the 16 frozen entries plus conditional ECC; not added to query recall |
| Inherited records outside current queries | 3 | Prior proof retained rather than silently deleting them |
| Full register | 1,610 | 520 accepted, 618 deferred after review, 424 rejected/out of scope, 48 already recorded |
See paper-register.json for reasons, full-text SHA, reading locations, and inherited sources. Acceptance adds a paper to the suitable implementation queue; it does not mean full original-model/environment reproduction. Implementation reconciliation independently checks catalog, executable registration, and GPU evidence. It does not infer capability gains from acceptance. See batch scope for acceptance of the 16 frozen implementations.
Full-text review evidence¶
- 12 industrial recommendation/search/ads items: method, production experiments, and limits recorded separately.
- 105 post-training items: public objectives and evaluation boundaries. Google LES was explicitly deferred, not rejected for missing abstract A/B language.
- 139 primary foundation-model items and four prior cross-checks, plus the other 119.
- 203 Agent items: 203 abstracts, 198 bounded full-text checks, five unavailable full texts explicitly deferred.
- Independent frozen shortlist sources: four post-training papers, six Agent/Jev papers, and five foundation/multimodal papers. FORTE is included in the 12 industrial entries.
- Conditional ECC: real online click evidence confirmed, but private logs, unpublished code, and a commercial image backend defer implementation outside the current 16.
Reading scope is marked per entry: decisive methods and experiment passages, not every paragraph and appendix. Code links are evidence to verify, not proof that a dependency or web shell is a complete author implementation.
Official institutional sources¶
| Source | Catalog scope | Actual verification |
|---|---|---|
| Google Research | 28 pages, 406 entries | Inherited proof for 398 URLs; individually checked eight genuinely new URLs for identity and date; excluded Preview/JavaScript noise |
| Meta | Six pages, 144 entries | Identical prior URL set; refreshed pagination and inherited decisions without recounting new papers |
| Google DeepMind | Nine pages, 265 entries | Official dates checked individually and predate the window; not 265 full algorithm rereads |
Receipts: Google, Meta, DeepMind, and Google new-entry verification. Google/Meta arXiv affiliations were also traced from manuscripts, not institutional catalogs alone. Production experiments in the body take priority; omission of A/B from an abstract is not a rejection reason.
The 33 archived exceptions are non-blocking and reopen only with new evidence. Do not archive confirmed in-window unread candidates to evade review. The strict source-exhaustion watermark is retired; all_sources_exhaustive=false and implementation_complete=false remain unchanged.
Follow-up candidates¶
Keep individual rationale for new qualifying leads separate from the 16 accepted implementation commitments. Google/Meta Self-Preservation, MetaEncoder, and X-AgentRed, and industrial LLM-BlockFE need distinct data, evaluation, or safe execution boundaries. Verify by priority in later batches; do not advertise unimplemented candidates as capability implementations.