October 25, 2026: Meta OPD, Inside-JEPA, and Google historical backfill¶
Three low-cost L1 components passed frozen-source CPU and real single-A100 acceptance over three seeds. Two are same-announcement discoveries and one is historical, not three October 25 first publications. The full large-scale experiments are too costly, not reproduced (review DONE) and do not enter retry or implementation queues.
Deliverables and limits¶
| Paper | Implemented | Not reproduced |
|---|---|---|
| Meta ME-OPD (P0) | Strict edit masks under a frozen old policy, correct/incorrect branches, and selected-token normalization over the full minibatch; undisclosed task objective explicitly supplied externally | No VLM, teacher, or verifier execution, reference description/solution generation, or math/vision capability claims |
| Inside-JEPA (P1) | Ordered relation states, online-future inverse gradients, EMA 0.996, five-arm controls, and actual AdamW updates | No video backbone, detection/tracking, SSv2/RH20T, or planning evaluation; undisclosed head topology supplied by caller |
| GroupDPO (Google/DeepMind collaboration, historical) | Four exact group objectives, full-group coefficients, positive NLL, and microbatch backward passes; multistep SGD/AdamW gradient equivalence | No 4B–32B LLM, reward model, or online sampling; no measured end-to-end memory or latency claim |
Paper pages include exact v1 times, primary first-author affiliations, author release status, stable keys, code paths, and figure sources and SHA-256. No verified author code means “not found,” not “definitely unreleased.” GroupDPO's canonical identity remains v1 on 2026-04-17T00:56:59Z; local code follows September 1 v2, reported separately from first publication.
Execution¶
Final frozen commit: cc8126d359c51676920f8a072b5a46374254e13e.
- CPU raw results: three components × seeds 42/43/44, about 1.40 seconds.
- A100 raw results: same three components and seeds, about 2.30 seconds.
- Frozen-source proof binds source hashes, commit, and paths for offline shallow-CI checks.
- Per-paper
metrics/component-cuda-seeds42-44.jsonfiles and three sanitized receipts indocs/gpu-validations/remain separate.
Timings exclude connection, syncing, and tests and are not full training times. Cross-review found extreme continuous Inside-JEPA controls could overflow squared terms or total loss; explicit rejection and three negative tests were added before rerunning CPU/A100 from the fixed commit. Numerical mechanisms, real gradient updates, and negative controls do not establish pretrained capabilities, formal benchmarks, or evolve promotion.
python scripts/build_oct25_component_artifacts.py --check
python -m pytest tests/test_oct25_component_evidence.py
Acceptance binds exact checks/metrics, bounds, executed device, seeds, frozen source, raw bytes, catalog, per-paper metrics, and scope/ledger identities. Relabeling CPU as CUDA, wrong dates, unaccepted components, and altered promotion flags fail. Cost closure is distinct from acceptance.
Discovery counts¶
Source coverage: eight arXiv pages still dated 2026-10-23, 1,260 category records and 916 identities. Version lists and ordering match the prior batch, with zero new identity or version strings; cross-listings, replacements, and first discovery are not new publications. ME-OPD and Inside-JEPA are same-announcement follow-ups; GroupDPO came from a historical conference listing.
Google and Meta catalogs timed out, giving unknown/null, not zero. Nine DeepMind pages and 265 records yielded no bounded differences, not exhaustive coverage. Independent priority review records the successful official Google EMNLP 2026 page: 31 means listed entries, not new papers, and curl did not retrieve raw HTML, so no hash was invented.
Title and identity audit splits 31 entries into two already listed (including GroupDPO), 25 historical metadata leads (19 from 2026 and six from 2025), and four failed-link or mismatched exceptions. These leads had no full-text review, qualification, cost evaluation, or implementation acceptance in this batch and were not queued.
This snapshot remains historical. The later approved closure of all 25 delivered 14 components, eight high-cost not reproduced, and three unaccepted; do not continue counting the old leads as pending.
Other targeted findings in additional_dispositions:
- Meta GenIA: v1 October 8; changed push time did not change October 9 release HEAD. SAM3D plus geometry, rendering, and test-time optimization were too costly and not reproduced (review DONE); no checkpoint or rendering stack was acquired.
- GRIP sparse retrieval: 30 vCPUs, 222 GB RAM, 8.8M–138M passages, and GB-scale indexes; too costly, not reproduced, review DONE. No term-seed heuristic substituted for SparsePiPNN and graph search; distinct from post-training GRIP.
- Meta AmbiGUI: official source release verified. The README awaits a citation after publication; no canonical paper identity or first date yet. Record code release only, not a fabricated paper, capability integration, or implementation promise.
Global review remains 2026-10-15. Three fixed deliveries do not close October 16–25, conference listings, or full texts. Inaccessible archived methods were not reopened periodically.