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.

{
    "construct": "evidential-tags-obs-inf-rep-src-with-instrument-recall-and-p-2",
    "metric": "token_delta",
    "formula_version": 1,
    "models": [
        "tiktoken/[email protected]",
        "tiktoken/[email protected]"
    ],
    "test_set": [
        {
            "english": "I directly observed that the ballot closed at quorum.",
            "ainglish": "obs: the ballot closed at quorum."
        },
        {
            "english": "I directly observed that the anchor row flipped to confirmed.",
            "ainglish": "obs: the anchor row flipped to confirmed."
        },
        {
            "english": "I directly observed that the second was recorded against the successor row.",
            "ainglish": "obs: the second was recorded against the successor row."
        },
        {
            "english": "The registry health endpoint reported that all migrations are applied.",
            "ainglish": "obs(health endpoint): all migrations are applied."
        },
        {
            "english": "The coverage gate's output reported that the floor was met.",
            "ainglish": "obs(coverage gate): the floor was met."
        },
        {
            "english": "The wire check reported that twenty-nine anchors are present.",
            "ainglish": "obs(wire check): twenty-nine anchors are present."
        },
        {
            "english": "I deduce, from premises I have not stated, that the seconder read the diff.",
            "ainglish": "inf: the seconder read the diff."
        },
        {
            "english": "I deduce, from premises I have not stated, that the queue stalled overnight.",
            "ainglish": "inf: the queue stalled overnight."
        },
        {
            "english": "From the vote timestamps and the ratification record, I infer the closure rule fired; the claim stands no stronger than the weaker of those two sources.",
            "ainglish": "inf(vote timestamps + ratification record): the closure rule fired."
        },
        {
            "english": "From the manifest hash alone, I infer the inputs were disjoint; the claim is bounded by that single source.",
            "ainglish": "inf(manifest hash): the inputs were disjoint."
        },
        {
            "english": "From the changelog entry and the register diff, I infer the slot was renumbered; the claim stands no stronger than the weaker of those two sources.",
            "ainglish": "inf(changelog entry + register diff): the slot was renumbered."
        },
        {
            "english": "According to the proposer's filing, the panel design is frozen.",
            "ainglish": "rep(proposer's filing): the panel design is frozen."
        },
        {
            "english": "According to the operator map, the two accounts share one principal.",
            "ainglish": "rep(operator map): the two accounts share one principal."
        },
        {
            "english": "According to the suggestions feed, the row is one second short.",
            "ainglish": "rep(suggestions feed): the row is one second short."
        },
        {
            "english": "Recalled from my own earlier state and unverified now: the gate never fired.",
            "ainglish": "rep(self-past): the gate never fired."
        },
        {
            "english": "Recalled from my own earlier state and unverified now: the digest matched at close.",
            "ainglish": "rep(self-past): the digest matched at close."
        }
    ],
    "tokenizers": "tiktoken 0.13.0",
    "method": "For each pair and tokenizer, delta = len(encode(ainglish)) - len(encode(english)) with tiktoken cl100k_base and o200k_base (versions pinned in roster names); per-tokenizer value is the mean over sixteen pairs; headline value is the MAX of the two per-tokenizer means (lower_better worst case), mirroring the replicated original's aggregation; lo/hi are min/max per-pair deltas across both tokenizers. Replication of 2cf05685... with disjoint metric inputs: all sixteen pairs are novel, agent-governance/measurement domain, no pair byte-matches the original's. English sides render the construct's full declared semantics from english_mapping, including the weakest-premise bound for inf(premises) pairs.",
    "form_counts": {
        "obs": 3,
        "obs(instrument)": 3,
        "inf": 2,
        "inf(premises)": 3,
        "rep(src)": 3,
        "rep(self-past)": 2
    },
    "seed": "none - deterministic, no sampling"
}