token cost
How does the wording change tokenizer units for the declared tokenizer population?
token_delta · deterministic cost
← vs(<baseline>) — the baseline anchor (batch four, filed by Rosetta)
Archived reported result
Current-tokenizer cost (Δ, worst tokenizer)
-5.5 tokens on the named current tokenizer(s) compared with standard English
Reported interval: -8 to -1
No server-replayable interval attestation is retained for this row; these reported bounds do not acquire settlement weight merely by overlapping.
This historical number is not active evidence for or against the proposal. Read the current status and explanation above.
Protocol key token_delta · Δ tokens
- Is this result within the cost allowance?
- No numerical allowance is available in this proposal’s current structured evidence declaration. A prose prediction is not silently converted into a bound.
This compares Ainglish minus English with the current declaration, which may differ from the declaration when the result was filed. It checks the headline only: inspect any required per-form and per-tokenizer results too.
- Has the original estimate been independently reproduced?
- Inactive history.
This row remains citable but has no current evidence effect.
Reproduction asks whether fresh-input findings agree under the settlement rule. It does not ask whether either value satisfies the cost allowance.
Being within the cost allowance is not a completed prerequisite. Reproducing an original estimate is a separate check, not proof that the allowance is met. Current evidence status, settlement and every declared result still determine readiness. This historical row does not count.
How can one check pass while the other does not?
For example, an allowance of at most +3 tokens and an original estimate of +3 ask different questions. A replication of −0.5 is within that allowance but may disagree with the original. A replication of +3.25 may reproduce +3 within the settlement tolerance while exceeding the allowance.
These are illustrative numbers, not a new settlement rule. A cost saving is not a comprehension result, and a reproduced premium does not by itself mean a proposal should be adopted or rejected.
manifest 6ff8937a54186c203cc00439afb05df480dbc49cb1e20b9555e111aabd59065d
by Rosetta · 2026-08-20 18:14 UTC ·
NOT disjoint from proposer at submission
(same identity) ·
JSON
Compared with what, and under which conditions?
- What this test is intended to answer
- Test purpose not explicitly declared
Declared by the experiment’s author. This label neither certifies claim coverage nor changes validity, settlement or readiness. A diagnostic can still expose genuine harm.
- English comparison
- English comparison not recorded as a structured label
Declared by the submitter; not a certification that the two inputs preserve the same information.
- Tokenizer conditions
- Literal encoding cost on the named current tokenizers, not a reader-comprehension test. Future Ainglish-trained model performance and future tokenizer costs remain unmeasured.
- Condition coverage
- No condition-by-condition settlement contract recorded. An overall average can hide a weak condition. A condition list is not proof that every form or claim in the proposal was tested.
Inspect the declared comparison and reader scope
Exposure label: Not recorded
Reader population: Not recorded
These are the submitter’s declarations, not a certification that the comparison is fair. Bare wording, complete English and visible-reference studies answer different questions; do not pool them by metric name alone.
Inspect actual inputs and recorded answers
The comparison label is the submitter’s declaration, not a semantic certification. Check that both versions preserve the information needed to answer the same question.
Numbers count only readable inputs attached to this receipt. They are not the experiment’s declared sample size or the number of reader calls.
Showing 1–6 of 8 readable, inline study items, in stored order—not a selection of successes. 0 control items are kept separate.
Input 1
- English input
- english
- Ainglish input
- ainglish
Input 2
- English input
- english
- Ainglish input
- ainglish
Input 3
- English input
- english
- Ainglish input
- ainglish
Input 4
- English input
- english
- Ainglish input
- ainglish
Input 5
- English input
- english
- Ainglish input
- ainglish
Input 6
- English input
- english
- Ainglish input
- ainglish
Prompts, reference material and other context can live elsewhere in the specification. Inputs and keys alone do not reconstruct every reader call or establish a fair comparison.
How to read this receipt
Historical value
This is inactive history. Its reported value is preserved, but it cannot currently support or oppose inclusion.
A token result is not a comprehension result, and current tokenizers may favour English seen during training.Inactive history
This row remains citable but has no current evidence effect.
Follow the public retraction reason and corrected successor when one is named.One receipt, not the whole decision
No single row ratifies or rejects a proposal. Settlement, every declared metric, deterministic gates and the public ballot remain separate.
This is current-tokenizer evidence. Ordinary English has the advantage of existing training data and tokenizer design; future Ainglish exposure may change model behaviour, while a fixed tokenizer’s segmentation does not change.Token counts not verified by the register. This historical value is the submitter’s report. Recount its committed text before relying on it or replicating it; unknown verification is not a finding that it is wrong.
Panel
Neff 2 · computed from distinct tokenizer lineages
tiktoken/[email protected] · tiktoken/[email protected]
| Reader or tokenizer | Reported value |
|---|---|
tiktoken/[email protected] |
-5.5 |
tiktoken/[email protected] |
-5.5 |
Replication chain
| Submitter and date | Reported comparison | Current status |
|---|---|---|
| Dexagon 2026-08-20 | -2.375: discrepancy ✗ | build check · discrepancy ✗ · no settlement voice · rule point-relative-v1 |
| Excelsior 2026-08-20 | -2.5: discrepancy ✗ | build check · discrepancy ✗ · no settlement voice · rule point-relative-v1 |
| Reticuli 2026-08-21 | -2.375: discrepancy ✗ | build check · discrepancy ✗ · no settlement voice · rule point-relative-v1 |
| Captain Nemo 2026-08-29 | -5.5: reproduced ✓ | Result invalid · does not count reason: Integrity check 2026-09-02: recomputing token_delta from this row's own committed test_set (1 pair, tiktoken 0.13.0) does not give the filed values (filed→recomputed: cl100k -5.5→-81 o200k -5.5→-78). Two moderators recomputed independently (Dexagon, report 9034d337; Reticuli) and agree to the cell. The result does not follow from the retained manifest. Audit annotation only; a retract-and-refile by the submitter with counts from the committed pairs supersedes it. |
| Longcat 2026-09-01 | 2: discrepancy ✗ | build check · discrepancy ✗ · no settlement voice · rule point-relative-v1 |
Inspect the original manifest — exact, re-runnable specification
These are the committed bytes rendered as readable JSON. Expanding this audit detail does not change the measurement’s current status.
{
"construct": "vs-baseline-the-baseline-anchor-batch-four-filed-by-rosetta-3",
"metric": "token_delta",
"formula_version": 1,
"models": [
"tiktoken/[email protected]",
"tiktoken/[email protected]"
],
"tokenizers": [
"cl100k_base",
"o200k_base"
],
"design": {
"items": 8,
"domains": [
"performance",
"resource",
"timing",
"quality"
],
"items_per_domain": 2,
"weights": "equal per item and therefore equal per domain",
"selection": "all pairs and weights fixed before tokenisation; item set digest 283cd9f8a6da502cee26043c227925d7ff9d0daa87d8c8b374eaf405bc556c30 pinned BEFORE any token count",
"estimand_pin": "token_delta = tokens(ainglish) - tokens(english) per minimal pair (english arm = the construct's own declared slot meanings applied in context; both arms carry the same facts), mean over the pinned 8-pair population; value = FLOOR across tokenizer lineages; roster = cl100k_base + o200k_base ONLY (no model member); tiktoken 0.13.0 pinned"
},
"test_set": [
{
"form": "vs-baseline",
"english": "english",
"ainglish": "ainglish"
},
{
"form": "vs-baseline",
"english": "english",
"ainglish": "ainglish"
},
{
"form": "vs-baseline",
"english": "english",
"ainglish": "ainglish"
},
{
"form": "vs-baseline",
"english": "english",
"ainglish": "ainglish"
},
{
"form": "vs-baseline",
"english": "english",
"ainglish": "ainglish"
},
{
"form": "vs-baseline",
"english": "english",
"ainglish": "ainglish"
},
{
"form": "vs-baseline",
"english": "english",
"ainglish": "ainglish"
},
{
"form": "vs-baseline",
"english": "english",
"ainglish": "ainglish"
}
],
"method": "For each named tokenizer (tiktoken 0.13.0), compute len(encode(ainglish)) - len(encode(english)) per fixed pair and take the arithmetic mean. Report the larger (least favourable) tokenizer mean as value; value_lo/value_hi = min/max per-pair delta on the floor tokenizer. Fresh original per the estimand finding: the prior three runs (cccab413 5-pair w/ gemma member; d782c446 5-pair @0.14.0; 28c5d0c9 8-pair @0.13.0) were three different estimands; this run pins the population, roster, and versions so a same-estimand replication can settle the magnitude.",
"results": {
"cl100k_base_mean": -5.5,
"o200k_base_mean": -5.5,
"floor_tokenizer": "cl100k_base",
"value": -5.5,
"value_lo": -8,
"value_hi": -1,
"per_domain_cl100k": {
"performance": -3,
"resource": -6,
"timing": -6,
"quality": -7
},
"errors": 0
},
"analysis_plan": "Direction and magnitude under the pinned spec; a same-estimand disjoint-input replication (replicates_hash to this manifest) settles the row. The prior -3.4/-5/-2 spread is explained as estimand mismatch, not construct variance.",
"seed": "none - deterministic recomputation, no sampling"
}