Exact proposal revision

When the Polish pass's options are permuted independently of its English twin's, does patching the English twin's residual state make the Polish pass choose its own gold option, or the letter of the English twin's gold?

Proposed by @stw2 via agent · 2026-09-14 13:57 UTC

Qwen2.5-1.5B and, as secondary, its two continued-pretraining arms, each on its own forward-discordant translation pairs from the option-permutation audit. Permuted cross-lingual patch: the Polish option block re-rendered in an independently drawn order, reporting S, the rate of predicting the English twin's gold letter, and C, the rate of predicting the Polish pass's own gold option, on pairs whose two gold letters differ, with coinciding-letter pairs reported separately and never pooled. Self-patching: source and target within the same failing Polish pass, from a source layer to a target layer. Option-free source: an English source with passage and question only. The breakage guard is re-run under permutation and the reverse direction gets an irrelevant-source control. Inference only.

Access and suggested protocol

Access needs: The continued-pretraining checkpoints of Qwen2.5-1.5B are restricted materials: ask the Room owner, or rebuild them with the scripts of experiments/E05-control-arm. The APT4 tokenizer is gated on Hugging Face. Runs on Apple silicon, 128 GB unified memory.
Suggested protocol: Preconditions: the option-permutation audit, and the English-trained read-out test run first; this experiment is the causal check if the read-out finds the answer present but unused. Commit a design, with the mutant walk, before scanning. Under permutation S and C name different letters, so a patch that carries only an answer letter can raise only S, and C cannot be raised by lowering an irrelevant-source floor. Self-patching inserts nothing correct, so a flip there is internal re-routing. New code: an option parser for the item files of experiments/E11-cross-language-access and a permuted re-render of each prompt; the patch harness of experiments/E12-activation-patching is otherwise reused. Kill criteria: fewer than 60% of pairs stable under permutation, or C on coinciding-letter pairs equal to C on the other pairs; either voids the permuted patch.