The missing control arm

Proposed open work.

Proposed open work.

Proposed open work.

Proposed open work.

At matched 0.5B-token continued pretraining of Qwen2.5-1.5B, continued pretraining is 0.9482 of the Polish bits-per-byte movement, while the APT4 transplant arm stays 0.12 to 0.25 bits per byte above the original-tokenizer arm on English and formal texts. The 1B extension was not launched.

Knowledge probe near chance for both arms; it bounds, but cannot resolve, capability differences.

Knowledge probe near chance for both arms; it bounds, but cannot resolve, capability differences.

A tokenizer-specific arithmetic deficit remains after matched continued pretraining, beside a larger decline from continued pretraining alone. Direction: both ways; inconclusive, as the design fixed no decision rule.

Continued pretraining alone accounts for a large part of the arithmetic decline of the transplant pipeline. Direction: as a substantial continued-pretraining share predicts; inconclusive, as the design fixed no decision rule.

The transplant's arithmetic before any recovery training.

Why sci_gsm8k is excluded from inference.

Whether the transplant residual persists or closes with more training is not decided at 0.5B tokens.

Whether the transplant residual persists or closes with more training is not decided at 0.5B tokens.

Whether the transplant residual persists or closes with more training is not decided at 0.5B tokens.

Whether the transplant residual persists or closes with more training is not decided at 0.5B tokens.

Whether the transplant residual persists or closes with more training is not decided at 0.5B tokens.

Whether the transplant residual persists or closes with more training is not decided at 0.5B tokens.

Continued pretraining accounts for most of the Polish movement at matched budget. Direction: as a substantial continued-pretraining share predicts; inconclusive, as the design fixed no decision rule.

Contamination-decay reference, excluded from inference by the note of 5 July 2026.

Continued pretraining alone moves this text little, and in the other direction. Direction: against a substantial continued-pretraining share; inconclusive, as the design fixed no decision rule.

Continued pretraining alone moves this text little, and in the other direction. Direction: against a substantial continued-pretraining share; inconclusive, as the design fixed no decision rule.

Continued pretraining alone moves this text little, and in the other direction. Direction: against a substantial continued-pretraining share; inconclusive, as the design fixed no decision rule.

Continued pretraining alone moves this text little, and in the other direction. Direction: against a substantial continued-pretraining share; inconclusive, as the design fixed no decision rule.

Continued pretraining alone moves this text little, and in the other direction. Direction: against a substantial continued-pretraining share; inconclusive, as the design fixed no decision rule.