Sixteen Google-priority items: low-cost component delivery¶
This round implements exactly 16 accepted follow-ups from the previous 39-item review, not 16 new publications or full reproductions. The historical review snapshot is unchanged; a new ledger batch records component delivery.
Delivered scope¶
| Paper / component | Actual delivery | Verification boundary |
|---|---|---|
| Dynamic Cogeneration | BRT original defect failure, patched pass, and grouped patch selection | Real subprocesses, manual patch; no LLM repair |
| ME-POIs | Multiscale Gaussian prior and contrastive/KL/text objectives | A100 tensor and gradient fixture; no private mobility data |
| Attribution Sets | v1 formula-level risk estimation and finite-candidate ERM | Retained finite-sample counterexample to printed formula; unbiasedness not verified |
| CausalArmor | LOO likelihood attribution and masking/regeneration interface | Real 0.5B likelihood on A100; no safety accuracy |
| Spatial Ads | Grouped pseudo-disk LP and randomized rounding | CPU math subproblem, not full matching/payment/PTAS |
| ConvApparel | PLSA statistics, HLS, and cross-policy validation | CPU diagnostic classifier is not Gemini; no human data collection |
| Modular Robust Generative | Globally constrained primal-dual, SIR, and router objectives | A100 finite support; residual did not converge in 16 steps; not robustness capability |
| Data Quality LSA | LSA initialization, rollout, closed-form SFT, and OS objective | A100 formulas and gradients; OS is regression, not policy-gradient RL |
| POLCA | Epsilon-net and search history containing inputs and outputs | CPU callback fixtures; no LLM program optimization |
| Simula | Taxonomy, complexity expansion, independent dual critics, and complexity protocol | Injectable model interface with zero model calls; not a data quality result |
| TFRBench | MASE, reference construction, and isolated held-out predictions | Reference construction reads labels and is not leakage-free capability evidence |
| ProEval | Empirical GP/BQ and failure-set capture | CPU historical score fixture; no rerun of LLM evaluation |
| mild | Margin logistic objective, gradients, and router | CPU mathematical training; inference does not read answers/costs |
| ScientistOne | CoE evidence binding, score rechecks, CPR, and audit interface | Missing semantic/citation evidence stays unverified |
| Activation Model Scanner | 48 official input pairs, activation directions, layer scans, and structural comparison | 96 real A100 forward passes; no independent baseline or safety certification |
| HCRG | Python AST graph/skeleton, three hops, dependency order, and static metrics | Not tree-sitter plus Gemini Java migration |
The scope manifest records evidence, full-experiment status, and costs. component_only and evaluation_protocol_interface are not full_pipeline. All are diagnostic_only=true, with no capability gain or evolve integration.
Full experiments¶
Full multi-data, multi-model, paid-API, multi-seed, or private-data experiments are not reproduced (review DONE) and not automatically retried. ME-POIs documents its unavailable private data. Spatial Ads is theoretical, with no full training experiment; do not invent a cost decision to fill the table. These statuses are independent of component delivery.
Cross-review fixes¶
- HCRG stops guessing unimported short names across modules and keeps parameter/local shadows unresolved.
- POLCA history includes inputs and program outputs, not just scores and text.
- AMS reads
hidden_states[block+1]for official zero-indexed blocks, verified by a real small Transformer hook and rerun on A100. - Robust gating separates theoretical KL from practical adversary NLL; symmetric entropy fixtures must not hide the distinction.
- Attribution Sets keeps the v1 printed formula and independent finite-sample counterexample, without silently rewriting the equation to pass.
Early/intermediate runs stay separate from final acceptance. Receipt status “passed” covers only specified mechanisms, execution, and provenance, not capability, safety, or optimality.
Checks¶
PYTHONPATH=src python scripts/closeout_google16_components.py --check
python -m pytest tests/test_google16_*.py
python scripts/validate_gpu_evidence.py
The earlier table saying 16 items were unimplemented remains the review snapshot from that date. This round did not run a new cross-domain scan or advance the 2026-10-15 watermark.