Current-tokenizer cost (Δ, worst tokenizer)

These are the committed bytes rendered as readable JSON. Expanding this audit detail does not change the measurement’s current status.

{
    "metric": "token_delta",
    "construct": "unless-the-plain-english-falsifier-claim-tag-in-words",
    "models": [
        "cl100k_base",
        "o200k_base",
        "google/gemma-4-31b-it"
    ],
    "test_set": [
        {
            "english": "The deploy is green — that claim fails if the smoke suite lied.",
            "ainglish": "the deploy is green unless(the smoke suite lied)."
        },
        {
            "english": "The cache is warm — that claim fails if the TTL was misread.",
            "ainglish": "the cache is warm unless(the TTL was misread)."
        },
        {
            "english": "The backup is complete — that claim fails if the manifest undercounts.",
            "ainglish": "the backup is complete unless(the manifest undercounts)."
        },
        {
            "english": "The quorum was met — that claim fails if a vote was double-counted.",
            "ainglish": "the quorum was met unless(a vote was double-counted)."
        },
        {
            "english": "The mirror is current — that claim fails if the cron silently died.",
            "ainglish": "the mirror is current unless(the cron silently died)."
        },
        {
            "english": "The row is settled — that claim fails if the two manifests secretly differ.",
            "ainglish": "the row is settled unless(the two manifests secretly differ)."
        }
    ],
    "seed": "deterministic — tokenizer counting involves no sampling",
    "prompts": "none — arms tokenized directly (tiktoken get_encoding().encode; transformers AutoTokenizer.encode add_special_tokens=False)",
    "method": "delta = tokens(ainglish) - tokens(english) per pair; member value = mean; value = least favorable member (max). Six pairs; english arms carry the construct's FULL payload — the claim plus the falsifier attached as part of the claim ('that claim fails if F'), compact phrasing — because plain-English 'unless' does not pin falsifier semantics (it usually reads as a conditional exception), so an english arm using bare 'unless' would under-translate the construct and flatter the delta.",
    "instrument_versions": {
        "tiktoken": "0.13.0",
        "transformers": "5.14.1"
    },
    "per_pair": {
        "cl100k_base": [
            -4,
            -4,
            -4,
            -4,
            -4,
            -4
        ],
        "o200k_base": [
            -4,
            -4,
            -4,
            -4,
            -4,
            -4
        ],
        "google/gemma-4-31b-it": [
            -3,
            -3,
            -3,
            -3,
            -3,
            -3
        ]
    },
    "reading": "Uniform -4/-4/-3 across six pairs and three lineages; value -3.0 (least favorable). The saving is real but modest — the construct's primary claim is the claim-tag's falsifier discipline in word-carried form, not compression; this row prices the surface so the comprehension question (does unless(F) read as falsifier-attached rather than conditional-exception?) is the one left open for a panel."
}