Hardening against October 2026 practice: what the study already leads on, and Plan H¶

Five literature sweeps (inference and design; evaluation and data; contamination; federated LoRA privacy; conventions and fine-tuning), each paper opened on arXiv, DOI or the venue page; most were read at abstract level, the full text where a claim below rests on it. Decided as ROADMAP Plan H. Claims about this repository were checked against the code before being written here.

Already at or beyond current practice. The own-half, sibling-half and foreign contrast has no counterpart: industrial studies (MetaMateCR, arXiv:2507.13499) compare a fine-tuned model with a zero-shot one and never separate internal APIs from convention. Averaging estimates over partitions with one crossed interval beats the median-over-splits rule (Chernozhukov et al., Econometrica 2025) and p-value merging (Gasparin, Wang and Ramdas, PNAS 2025); the SESOI, bounded reading, outcome-neutral checks and seal go beyond the SE registered-report template (arXiv:2602.09292). Deterministic, source-read bot rules beat LLM filtering of review data (Too Noisy To Learn, MSR 2025, arXiv:2502.02757). The model-less baseline in the contamination battery is what the MIA literature asks for (arXiv:2406.16201, arXiv:2406.17975) and few SE studies run. The LoRA recipe (all attention and MLP projections, loss on the answer only, batch 16) matches LoRA Without Regret (Thinking Machines, 2025). No prior work attributes an organization through secure aggregation from LoRA updates, or ranks split designs by what their transmitted half leaks.

Checked here and confirmed. - Min-K%++ is the registered primary exposure measure (registered decisions, "Contamination"), and the battery's model-less bag-of-words baseline (balanced accuracy 0.589) beats it (AUC 0.540): the registration names an instrument this study's own data does not support. - No checkpoint is pinned by revision in sphragis/experiment/model.py. - lora_config ties alpha to 2r, so the scale is 2 at every rank and the rank-256 arm also takes larger steps; its readings do not isolate capacity. - CodeReviewer's exact match compares whitespace-collapsed strings (run_test_ref.py), so the pilot entry of 2026-09-14 compared its 30.32% with this study's strict 0.200; the like-for-like figure is the normalized 0.222. The conclusion there (exact match discriminates) stands. - registered-decisions.md described Holm as a fixed order; holm_steps implements the step-down. The wording now matches the code. - The registered test window (2025-11 to 2026-10) is the period in which AI-assisted commits rose fastest (agent-to-agent review grew more than 100-fold over 2025, arXiv:2608.21311), and nothing in the pipeline reads AI trailers on successor commits.

Sources, beyond those named above and in Plan H: Maini et al., NeurIPS 2024 (dataset inference); Hayes et al., NeurIPS 2025, arXiv:2505.18773; Cheng et al., Dated Data, COLM 2024, arXiv:2403.12958; Heineman et al., arXiv:2508.13144; Miller, arXiv:2411.00640; Bowyer et al., ICML 2025, arXiv:2503.01747; Liu, Hu, Wu and Smith, NeurIPS 2022, arXiv:2206.07902; Melis et al., S&P 2019, arXiv:1805.04049; Mitchell et al., arXiv:2606.10481; Athanasiou, Jung and Palamidessi, ICLR 2026, arXiv:2603.02017; Vijayvergiya et al., AutoCommenter, arXiv:2405.13565; Hora, Robbes and Zacchiroli, arXiv:2609.07542.