# Definition plus saved LoRA — frozen before inference, 2026-10-12

Four pinned base checkpoints and all five existing v2 adapters: 4B seeds17/29,
Qwen3.5-9B, Qwen3.5-27B and Qwen3.8-27B seed17. No retraining, optimizer or merge.
Load one adapter per fresh model. Record adapter file hashes, verify against previous runs.

Two matched conditions per adapter: original prompt (LoRA alone), and the exact
Italian definition from context_plan prepended to that prompt (definition plus LoRA).
No examples. Reuse all five holdouts and the same 2008 candidates, including EOS.
Reuse context-r2 scoring implementation with cached prefix, explicit positions,
batch8 sorted by length/index, BF16/SDPA and FP32 logsoftmax, TF32 off.
Repeat anchor batch after each case (bit-identical); compare cached singleton
against full singleton gold score. Preserve numerical discrepancies.

Generate greedy plus 3 samples seed17/29/43, T=.7, top_p=.8, top_k=20, max48tokens,
thinking disabled. Same exact-match normalization as the previous prompting probe.
Primary comparison: definition plus LoRA versus LoRA alone within this run.
Compare against prior base and definition-only context-r2 runs as frozen references.
Report rank, prefix rank, logP with/without EOS, recall@k, k50, all generations,
adapter identities and cache/full diagnostics. Do not treat a closed-pool rank as
an open-generation success, or five repeatedly observed cases as a new validation set.

Immutable run IDs: definition-lora-MODEL-sSEED. Store complete scores and source snapshots.