For the scaffold-final residual patch on 100 forward-discordant pairs of Qwen2.5-1.5B, 35 of the 44 odd-layer source and target pairs, among 196, whose transport rate is above 0.35 have a flip rate at least 3 times their irrelevant-source flip rate; the highest ratio is 5.8750, the highest flip rate 0.9800, and the mean flip rate 0.7557 against a mean irrelevant-source flip rate of 0.1784.

Relation: Count of layer pairs reaching a ratio

model

Qwen2.5-1.5BModel patched.

intervention

Cross-lingual residual patchChange applied.

anchor

Scaffold-final anchorToken position patched.

transport region

transport rate above 0.35Layer pairs summarised.

specificity ratio

3 ratioRatio of the pre-registered validity gate.

count reaching

35 layer pairsLayer pairs reaching the ratio.

count total

44 layer pairsLayer pairs in the region.

max ratio

5.8750 ratioHighest specific-to-irrelevant ratio in the region.

max flip rate

0.9800 ratioHighest flip rate in the region.

mean flip rate

0.7557 ratioMean flip rate over the region.

mean irrelevant flip rate

0.1784 ratioMean irrelevant-source flip rate over the region.

items

100 translation pairsForward-discordant pairs in the subsample.

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. Scan over the odd decoder layers 1 to 27 as source and target (196 pairs); region means are computed over the grid in results/analysis.json by scripts/06_metrics.py.

Dataset

Seeded subsample of 100 forward-discordant translation pairs of Qwen2.5-1.5B, 50 per benchmark.Version: unspecified · Access: public

Reported results

35/44 pairs reach ratio 3; max ratio 5.8750; max flip rate 0.9800; mean flip rate 0.7557, mean irrelevant 0.1784, mean transport 0.7436. Among the pairs reaching the ratio: mean flip rate 0.8494, irrelevant 0.1780, transport 0.8146.

Uncertainty and replication

Point values on 100 items per layer pair; no intervals in the scan.

Evidence references

scan_p1.jsonl · publichttps://github.com/stw2/tokenizer-science-tax/blob/633b3ba1ce916e432ae5ba4e0a706ad58b53b2e9/experiments/E12-activation-patching/results/scan_p1.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 irrelevant source usually carries a wrong letter for the target, which lowers the irrelevant-source flip rate and eases the ratio.