Compare two experiments

Compared with

Ainglish arm carrying checked(<predicate>@<checked-at>, scope=<scope>) minus English arm carrying the row's served lossless expansion 'I last checked <predicate> at <checked-at> against <scope>'; the predicate name, ISO timestamp and scope string are transcribed verbatim in both arms, so slot-filler cost cancels and the delta prices the construct's wrapper against the expansion's words (comparator genre: complete careful English, narrow checking-event reading, no result claim in either arm)

Tested population

16 fresh author-written messages: 8 standalone check reports and 8 reports embedded in a two-sentence operational message; 16 distinct predicate/timestamp/scope triples drawn from release, infrastructure and compliance contexts; no item reused from any prior row on this proposal

Unit tested

message pair: one short English message and its Ainglish counterpart, byte-identical outside the construct

How results combine

maximum tokenizer mean over the three declared tiktoken encodings (least-favourable); per-pair delta = Ainglish tokens minus English tokens; unstratified, so the headline is the plain item mean of the worst encoding

These are the study author’s declarations. A finding applies to this tested scope; this summary does not establish that another study is comparable.

Absolute arm results, reader-specific results and condition results below are retained values, not a newly pooled analysis. Accuracy arms use fractions from 0 to 1; their difference uses percentage points.

Absolute arm results

Not recorded

Reader or tokenizer results

[
    {
        "model": "cl100k_base",
        "value": -0.25
    },
    {
        "model": "o200k_base",
        "value": -0.3125
    },
    {
        "model": "p50k_base",
        "value": 1.75
    }
]

Condition results

Not recorded