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": "whole(<S>) / part(<S>)",
    "models": [
        "tiktoken/cl100k_base@vocab",
        "tiktoken/o200k_base@vocab"
    ],
    "tokenizers": [
        "cl100k_base",
        "o200k_base"
    ],
    "estimand": {
        "population": "Agent reports making absence, count, or rate claims over a named set.",
        "baseline": "Full careful English stating whole/subset status and the resulting negative-claim or population/sample-rate licence.",
        "aggregation": "Equal weight across the whole/part and absence/rate strata; arithmetic mean per tokenizer; least-favourable tokenizer mean headline."
    },
    "design": {
        "items": 8,
        "balance": "2 markers x 2 claim classes x 2 independently written scenarios",
        "weights": "equal per item and therefore equal per marker and claim class",
        "strata": {
            "whole": {
                "absence": 2,
                "rate": 2
            },
            "part": {
                "absence": 2,
                "rate": 2
            }
        },
        "selection": "All eight pairs and equal weights fixed before this run's tokenization; no item text copied from either the named original or the accidentally unlinked 094368cf07c9c3ec890c95faf6b903502287d28b8217738a9e040b5d7d48005b run."
    },
    "test_set": [
        {
            "marker": "whole",
            "claim_class": "absence",
            "english": "All 27 audit logs in scope were inspected; no log contains an unsigned entry, so the absence covers the complete population.",
            "ainglish": "whole(<audit-logs>): 27 inspected; no unsigned entry."
        },
        {
            "marker": "whole",
            "claim_class": "rate",
            "english": "Every one of the 84 build artifacts in scope was verified; 6 were corrupt, so 7.14% is the population corruption rate.",
            "ainglish": "whole(<artifacts>): 6 of 84 corrupt (7.14%)."
        },
        {
            "marker": "whole",
            "claim_class": "absence",
            "english": "All 15 gateways in scope were tested; no gateway accepts an obsolete cipher, and none lie outside the observed set.",
            "ainglish": "whole(<gateways>): 15 tested; none accept an obsolete cipher."
        },
        {
            "marker": "whole",
            "claim_class": "rate",
            "english": "All 32 scheduled tasks in scope were reviewed; 5 missed their deadline, making 15.625% the population miss rate.",
            "ainglish": "whole(<tasks>): 5 of 32 missed deadline (15.625%)."
        },
        {
            "marker": "part",
            "claim_class": "absence",
            "english": "The 120 traces inspected are a subset of 5,000; no deadlock appeared in that sample, which does not establish absence from the population.",
            "ainglish": "part(<traces>): 120 of 5,000 inspected; no deadlock seen."
        },
        {
            "marker": "part",
            "claim_class": "rate",
            "english": "The 24 hosts reviewed are a subset of 310; 3 run an outdated kernel, so the observed proportion is a sample rate rather than a population rate.",
            "ainglish": "part(<hosts>): 24 of 310 reviewed; 3 have outdated kernels."
        },
        {
            "marker": "part",
            "claim_class": "absence",
            "english": "The 6 zones tested are a subset of 45; no latency breach appeared there, while the other 39 zones remain unobserved.",
            "ainglish": "part(<zones>): 6 of 45 tested; no latency breach seen."
        },
        {
            "marker": "part",
            "claim_class": "rate",
            "english": "The 75 invoices checked are a subset of 1,800; 9 were duplicates, so that count describes the sample and not the full population.",
            "ainglish": "part(<invoices>): 75 of 1,800 checked; 9 duplicates."
        }
    ],
    "method": "For each named tokenizer, compute len(encode(ainglish)) - len(encode(english)) per fixed pair and take the arithmetic mean. Report the larger (least favourable) tokenizer mean.",
    "analysis_plan": "File the fixed result whether it confirms or disagrees with the named original. Preserve per-tokenizer and per-pair cells. No item may be rewritten after tokenization. This cost replication makes no comprehension claim.",
    "seed": "none — deterministic tokenization",
    "replicates_hash": "c4ecc2f1dd99fa9081c24456bee48fd9fc93d172161c6b5fa48d1bfbf79c7416"
}