The exploratory second model family is Llama-3.1-8B, not Qwen3-8B¶

The registered model's contamination argument is that the corpus (from 2024-10) postdates the checkpoint's publication. The exploratory second family, an H1 dev-window replication outside the registered cells (ROADMAP Plan F), was listed as Qwen3-8B; Alibaba states no cutoff for it, the model reports early 2025, which falls inside the training window, and it was published in 2025. A sweep of current open-weight families (October 2026) against two requirements, open weights (RQ1 trains adapters, RQ2 reads their updates) and a checkpoint published before the corpus:

• Hosted models are out on the first: Claude Haiku 4.5 cannot be fine-tuned, Claude 3 Haiku's Bedrock fine-tuning ended with the model (2026-09-10), and hosted tuning elsewhere returns an endpoint, not weights.
• Gemma 4 (April 2026; 2B, 4B, 26B MoE, 31B), gpt-oss-20b (August 2025, MoE in MXFP4) and the Granite 4 hybrids (mostly Mamba-2 layers) postdate the corpus or do not share the dense attention and MLP projections the registered LoRA configuration adapts.
• Llama-3.1-8B (Meta model card: pretraining data to December 2023, released 2024-07-23, Llama 3.1 Community License) meets both, at the registered size, with the same LoRA targets.
• Olmo 3 7B (Ai2, Apache 2.0, data to December 2024) overlaps the first two training months, but its training data is public, so exposure could be looked up rather than estimated; kept as the alternative if a reviewer presses on contamination.

Llama-3.1-8B-Instruct replaces Qwen3-8B in the plan; the membership probe reads its base checkpoint, as Min-K%++ reads Qwen2.5-Coder-7B's. Not yet run.