Recomputing Table 1 of arXiv:2604.10799v1 on the preamble's protocol text, without special tokens, with whitespace-split words and characters counted including whitespace, reproduces 60 of its 64 printed cells exactly; 7 of 8 rows match in every cell and the APT3 row in 4 of 8.

Relation: Reproduces

target

Table 1 of arXiv:2604.10799v1: tokens, characters per token and tokens per word of eight tokenizer rows on the Polish and English preamblePublished result recomputed.

protocol

protocol text variant; no special tokens; words split on whitespace; characters counted including whitespaceCounting protocol that matched best.

cells matched

60 cellsPrinted cells reproduced exactly.

cells total

64 cellsPrinted cells compared.

rows matched

7 rowsRows reproduced in every cell.

rows total

8 rowsRows compared.

Method and evaluation protocol

Tokenize the preamble texts with the nine tokenizers under every combination of text variant, special-token setting, word rule and character rule; count cells equal to the printed values; keep the protocol with the most matches.

Dataset

Preamble of the Constitution of the Republic of Poland, Polish text and official English translation, from Appendix A of the seed paper.Version: unspecified · Access: public

Reported results

60/64 cells; APT3 English tokens measured 618 against 615 printed; the shared Apertus and Mistral Small 3.1 row matches the Apertus tokenizer (547/377).

Uncertainty and replication

Exact integer counts; no sampling.

Evidence references

table1_replication.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/719d51eb4ac7e6e0f36d69cf339c8e976266ee98/experiments/E01-fertility-atlas/results/table1_replication.json

Limitations

Two text choices of the protocol (the Polish final sentence kept whole, the English ASCII apostrophe) were selected against the same 64 cells they are scored on. The preamble text was not cross-checked against the Sejm's published text.