# Pre-registration: does a CLAUDE.md cost the same in Chinese as in English?

Written 2026-08-27, BEFORE any run. Nothing below was revised afterwards.

## Why

On 2026-08-27 this site measured skill descriptions and found majority-Chinese ones cost about
2.2x per CHARACTER against English, with the arms not overlapping. Measured per BYTE of UTF-8 the
same arms OVERLAP. That says the per-character gap is largely a unit artefact, because a Chinese
character is roughly 2.3 UTF-8 bytes.

If that reading is right it makes a testable prediction somewhere else entirely. This site prices
CLAUDE.md per BYTE (106 tokens + 0.26 per byte). If the byte is the unit that travels, that rate
should hold for a Chinese CLAUDE.md without modification. If instead the tokeniser genuinely
penalises Chinese, the rate should come in low.

## The fixture

One translation pair, so meaning is held constant. Written for this test, not taken from a repo.

  claudemd-en.md   1,558 bytes   1,558 characters   ASCII
  claudemd-zh.md   1,536 bytes     556 characters   81% CJK

The bytes happen to match within 1.4% while the characters differ by 2.8x. That is what makes one
pair able to separate the two units.

## Predictions, all three recorded before running

  P1  This site's published CLAUDE.md rate, 106 + 0.26 per byte:
        EN ~511 tokens, ZH ~505 tokens. Near-identical.
  P2  The per-BYTE rates measured on skill descriptions the same day
      (English 0.363-0.482, Chinese 0.459-0.513 per byte):
        EN 566-751 tokens, ZH 705-788 tokens. ZH slightly HIGHER, ranges overlapping.
  P3  A naive reading of the per-CHARACTER headline, i.e. that Chinese is cheap because it uses
      fewer characters:
        ZH about 0.36x EN. ZH far cheaper.

P3 is the reading this site must not let a reader walk away with, which is why it is scored.

## How it is scored

Each arm is one CLAUDE.md at the project root of an otherwise empty project, one turn, measured
against an empty floor from the SAME round. Four rounds. The estimator is the guarded mode from
estimator.py, declared before the run.

A prediction is scored CORRECT if the measured mode falls inside its stated interval, WRONG if it
does not. P1 is a point prediction and is scored correct within +/-15%.

## Controls, all published per run

  PAIRED       floor re-measured every round; every delta is against that round's floor.
  ONE PATH     every arm runs in ONE constant directory, created and removed per run. This site
               has measured the startup floor moving with the working-directory path by up to
               about 80 tokens, which is large against this signal.
  ARRIVAL      the reply must be exactly OK.
  FIXTURE      the file's byte size is asserted at the project root before each run, and the
               floor arm asserts no CLAUDE.md exists.
  ISOLATION    --setting-sources project asserted onto the command line.
  529 GUARD    result_guard refuses a result whose modelUsage is empty, so an overloaded API
               cannot be recorded as a zero-token run.

## What this cannot show

One translation pair, written by the same author as the rest of this site's fixtures, and not
checked by a native speaker. It is one file at one size on one machine at one release. It says
nothing about a skill BODY, and nothing about any language outside these two.