The scaffold-final residual patch on Qwen2.5-1.5B fails the pre-registered validity gate at all 3 layer pairs confirmed on 326 forward-discordant pairs, with breakage of 0.1600 at 25 into 25, 0.1600 at 25 into 23, 0.7400 at 1 into 27 against at most 0.05, so the patch does not classify the English-minus-Polish accuracy gap as routing-bound or representation-bound.

Relation: Fails the validity gate

model

Qwen2.5-1.5BModel patched.

intervention

Cross-lingual residual patchChange applied.

anchor

Scaffold-final anchorToken position patched.

evaluation set

Forward-discordant translation pairsItems patched.

pairs failing

3 layer pairsConfirmed layer pairs failing the gate.

pairs confirmed

3 layer pairsLayer pairs confirmed.

specificity ratio

3 ratioRequired ratio of flip rate to irrelevant-source flip rate.

breakage limit

0.05 ratioHighest breakage the gate allows.

breakage 25 into 25

0.1600 ratioBreakage at source layer 25, target layer 25.

breakage 25 into 23

0.1600 ratioBreakage at source layer 25, target layer 23.

breakage 1 into 27

0.7400 ratioBreakage at source layer 1, target layer 27.

classification

routing-bound or representation-boundClassification the gate leaves undecided.

Method and evaluation protocol

The English twin's residual-stream state entering the source decoder layer at the scaffold-final token replaces, without scaling, the Polish twin's state entering the target layer while its four options are scored by length-normalised likelihood (torch MPS, bf16). Specific source: the item's own English twin. Irrelevant source: another item's English twin from a seeded within-benchmark derangement. Breakage: the same patch on a seeded guard of Polish-correct items. Validity gate as pre-registered: flip rate at least 3 times the irrelevant-source flip rate and breakage at most 5 percentage points.

Dataset

326 forward-discordant translation pairs of Qwen2.5-1.5B (151 Belebele and 175 MMLU pairs) from the Belebele and MMLU translation pairs; breakage guard of 100 Polish-correct pairs.Version: unspecified · Access: public

Reported results

25 into 25: ratio 4.0909, breakage 0.1600, transport 0.9632; 25 into 23: ratio 3.9744, breakage 0.1600, transport 0.9540; 1 into 27: ratio 0.9352, breakage 0.7400, transport 0.3098. Stored routing bins, voided by the gate: 25 into 25 ROUTING, 1 into 27 MIXED.

Uncertainty and replication

Breakage is a point value on 100 Polish-correct items; ratios use point estimates on 326 items.

Evidence references

scan_p2.jsonl · publichttps://github.com/stw2/tokenizer-science-tax/blob/633b3ba1ce916e432ae5ba4e0a706ad58b53b2e9/experiments/E12-activation-patching/results/scan_p2.jsonlanalysis.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/633b3ba1ce916e432ae5ba4e0a706ad58b53b2e9/experiments/E12-activation-patching/results/analysis.jsonmetrics.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/633b3ba1ce916e432ae5ba4e0a706ad58b53b2e9/experiments/E12-activation-patching/results/metrics.json

Limitations

One model family at 1.5B parameters. The source is the English twin, correct by construction of the forward-discordant set, and the endpoint is a single answer letter, so a late-layer patch can carry the answer itself. The forward-discordant set conditions on one scoring run; the letter priors of those scores were not audited. The 12-item transport probe that prompted the transport gate has no committed script or output and is unavailable. The confirmation pairs were chosen by a rule that differs from the pre-registered one (at most three by flip rate plus three transport-clean; pairs exactly 0.02 below the best excluded); the excluded pairs are transport-contaminated. results/analysis.json stores routing bins that the failed gate voids.