# Prompting controls — 2026-10-12
Frozen before inference. Four pinned original base checkpoints; no adapter and no training.
Three conditions: unchanged user prompt; 32 user/assistant worked examples followed by the same
user prompt; a general definition prepended to the same user prompt. No condition combines definition
and examples. All five previously confirmed holdouts, including the English clue, are evaluated.

Examples: first32 eligible training pairs ordered by SHA256('context32-v1:'+ID), no selection using
holdouts, no explanatory rationales. Definition authored for this experiment after consulting
https://www.gruppoludico.it/wiki-1/enigmistica/crittografie; no worked examples or holdout vocabulary.
Exact prompts/examples and candidate list are frozen in data/holdouts/context_plan.json.
This is 32-shot prompting versus LoRA trained on1592pairs for2epochs, not equal-data/equal-compute.

Candidate ranking uses the same2008canonical lowercase answers as PROBE_RANK.md. Rank sums token
logprobs with immediate EOS; answer-only score/rank also recorded. No length normalization, language
or diagram filtering. k50=third smallest of5ranks. English distractors mostly Italian; pool incomplete.
Greedy plus three samples per condition/case: seeds17/29/43,T.7,top-p.8,top-k20,max48tokens,no thinking.

Use BF16/SDPA with FP32 logsoftmax. Cache prefix except final prompt token; independently deep-copy
and replicate the pristine cache for each batch using reorder_cache(zeros), which supports both
attention and DeltaNet recurrent/conv states. Candidate batches of8 fixed by length then index.
Repeat an anchor batch after everycase and require bit-identical scores. Compare cached-singleton
gold against uncached-singleton gold for everycase; preserve errors without hiding discrepancies.
Cache smoke on9B: maximum per-token difference0.03824nat; summed difference0.03088nat.
Cached scoring may yield BF16 rounding differences vs the previous uncached ranks; rerun zero-shot
baseline with this same method. Existing LoRA ranks remain a separate frozen comparison.

Store all scores/generations, source/plan/protocol hashes, revisions, hardware, token counts and
numerical checks. No optimizer, no model updates, no adapter load, no change to training data.
Models4B/9B use A100-80GB;27B H100. Initial context-MODEL runs excluded: hybrid cached forward required explicit suffix position IDs.
Corrected runs context-r2-MODEL use explicit 2D suffix positions, expanded by the model to its text RoPE axes.
Runs context-r2-MODEL, immutable output directories.