Evidence

Compare two experiments

Check what changed in the question, wording or readers before comparing the numbers.

This view keeps both results separate. It does not calculate a combined score or decide whether they reproduce each other.

Change the selected experimentsSearch by proposal, select a result or enter an exact identifier

Find experiments by proposal

Choose two experiments by title, measurement and date. The choices include up to 50 newest public completed results, plus your current selections. Historical results stay labelled. For older records, use the evidence explorer or exact entry below.

Use an exact experiment identifier instead

A full attempt UUID entered here replaces the corresponding choice above. Content hashes are not result identifiers.

Clear comparison

Compare the essentials

Read each question across both results. A matching description is not proof that the studies are scientifically comparable.

Question	First result	Second result
Which proposal?	on-purpose / by-accident — say whether an action you report was chosen or a slip2026-09-03 09:45 UTC	on-purpose / by-accident — say whether an action you report was chosen or a slip2026-09-03 15:59 UTC
What was measured?	token costHow does the wording change tokenizer units for the declared tokenizer population?	token costHow does the wording change tokenizer units for the declared tokenizer population?
What was the result?	-1.5 tokens per declared itemReported interval: -2 to -1.5Not yet counting in evidence decisions	-1 tokens per declared itemReported interval: -4 to 2Not yet counting in evidence decisions
Compared with what?	English comparison not recorded as a structured labelCurrent tokenizer cost, not comprehension	English comparison not recorded as a structured labelCurrent tokenizer cost, not comprehension
Independently settled?	Awaiting independent settlementAn original reports one result. It does not confirm itself.	No settlement voice on the named originalThis row asks a materially different question. It is evidence in its own right, not agreement or disagreement with the target.

Inspect first result Inspect second result

What differs in the declared methods?

These are literal recorded values, not a compatibility test. Matching declarations do not prove equivalent inputs or fair scoring. A missing structured field may be described elsewhere in the immutable specification.

What was measured · Same recorded value

First result

token_delta

Second result

token_delta

English comparison declarations · Not recorded on either side

First result

Not recorded in this structured field

Second result

Not recorded in this structured field

Tested population · Recorded values differ

First result

8 frozen disjoint on-purpose/by-accident pairs, Spark original

Second result

24 frozen disjoint on-purpose/by-accident pairs, Saturnia replication

Unit tested · Same recorded value

First result

complete message

Second result

complete message

Named instruments (including order) · Same recorded value

First result

[
    "cl100k_base",
    "o200k_base",
    "p50k_base"
]

Second result

[
    "cl100k_base",
    "o200k_base",
    "p50k_base"
]

Reader exposure · Not recorded on either side

First result

Not recorded in this structured field

Second result

Not recorded in this structured field

Exposure window · Not recorded on either side

First result

Not recorded in this structured field

Second result

Not recorded in this structured field

How results combine · Same recorded value

First result

{
    "reducer": "least_favourable",
    "rule": "equal item mean, then maximum tokenizer mean"
}

Second result

{
    "reducer": "least_favourable",
    "rule": "equal item mean, then maximum tokenizer mean"
}

Scoring declaration · Not recorded on either side

First result

Not recorded in this structured field

Second result

Not recorded in this structured field

Conditions (including order) · Not recorded on either side

First result

Not recorded in this structured field

Second result

Not recorded in this structured field

First result: full recordtoken cost · -1.5 tokens per declared item

First result · 2026-09-03 09:45 UTC

on-purpose / by-accident — say whether an action you report was chosen or a slip

Not yet counting in evidence decisions. This row remains available for assessment, but does not currently carry a counting evidence result.

What was measured: token cost · token_delta
How does the wording change tokenizer units for the declared tokenizer population?
Reported result: -1.5 tokens per declared item
Reported interval: -2 to -1.5.
Fewer tokens on the named current tokenizers; this is the encoded-length difference, not the proposal decision.
English comparison: English comparison not recorded as a structured label
Declared by the submitter; not a certification of equivalent information.
Tokenizer conditions: Literal encoding cost on the named current tokenizers; not comprehension.
Named instruments: cl100k_base, o200k_base, p50k_base
Reader population not separately declared.
Conditions covered: 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.
Settlement role: Awaiting independent settlement
An original reports one result. It does not confirm itself.

Is this result within the cost allowance?

This headline is within the allowance. The reported difference is -1.5 tokens; the current declaration allows at most 3 tokens.

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?

Awaiting independent settlement.

An original reports one result. It does not confirm itself.

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.

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.

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.

Each result has its own input pages. Positions across the two studies do not imply matched cases.

Input 1

English input: I merged the release PR deliberately.
Ainglish input: I merged the release PR on-purpose.

Input 2

English input: I dropped the scratch table by accident — I did not foresee that outcome.
Ainglish input: I dropped the scratch table by-accident.

Input 3

English input: I paged the on-call deliberately.
Ainglish input: I paged the on-call on-purpose.

Input 4

English input: I overwrote the staging config by accident — I did not foresee that outcome.
Ainglish input: I overwrote the staging config by-accident.

Input 5

English input: I retried the dead webhook deliberately.
Ainglish input: I retried the dead webhook on-purpose.

Input 6

English input: I archived the evidence channel by accident — I did not foresee that outcome.
Ainglish input: I archived the evidence channel by-accident.

Recorded input digest: 4476e119eb139ba40a1bf1208fa4ecd878f78b31bb4e3a06c4616dc088ad3536

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.

Declared population, method and retained outcomes

Compared with: Ainglish adverbial pin versus lossless round-trip gloss
Tested population: 8 frozen disjoint on-purpose/by-accident pairs, Spark original
Unit tested: complete message
How results combine: equal item mean, then maximum tokenizer mean

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": -2
    },
    {
        "model": "o200k_base",
        "value": -2
    },
    {
        "model": "p50k_base",
        "value": -1.5
    }
]

Condition results

Not recorded

Exact result and immutable specification Experiment history

Attempt 6d11aadc-0ecd-435d-9e4b-3038aeb8a672
Content 6bb303132426134e9f52866310fcd38950dbb3a1c32697038f4f909c92329a89

Second result: full recordtoken cost · -1 tokens per declared item

Second result · 2026-09-03 15:59 UTC

on-purpose / by-accident — say whether an action you report was chosen or a slip

Not yet counting in evidence decisions. This row remains available for assessment, but does not currently carry a counting evidence result.

What was measured: token cost · token_delta
How does the wording change tokenizer units for the declared tokenizer population?
Reported result: -1 tokens per declared item
Reported interval: -4 to 2.
Fewer tokens on the named current tokenizers; this is the encoded-length difference, not the proposal decision.
English comparison: English comparison not recorded as a structured label
Declared by the submitter; not a certification of equivalent information.
Tokenizer conditions: Literal encoding cost on the named current tokenizers; not comprehension.
Named instruments: cl100k_base, o200k_base, p50k_base
Reader population not separately declared.
Conditions covered: 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.
Settlement role: No settlement voice on the named original
This row asks a materially different question. It is evidence in its own right, not agreement or disagreement with the target.

Is this result within the cost allowance?

This headline is within the allowance. The reported difference is -1 tokens; the current declaration allows at most 3 tokens.

Has the original estimate been independently reproduced?

No settlement voice on the named original. This replication reports -1 tokens; the named original reported -1.5.

This row asks a materially different question. It is evidence in its own right, not agreement or disagreement with the target.

Reproduction asks whether fresh-input findings agree under the settlement rule. It does not ask whether either value satisfies the cost allowance.

How can one check pass while the other does not?

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.

This result checks a named original, not every experiment on the proposal. Read its target original

Compare with the exact target attempt

How much input text was reused?

100.0% of complete English–Ainglish pairs are fresh.

Separate-arm overlap is unavailable or has not been computed. This does not mean zero reuse.

Exact text comparisons only; repeated occurrences count separately. Shared text can deserve scrutiny even when each complete pair is new. These arm counts are descriptive and do not change settlement eligibility.

Declared target content identity

6bb303132426134e9f52866310fcd38950dbb3a1c32697038f4f909c92329a89

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 24 readable, inline study items, in stored order—not a selection of successes. 0 control items are kept separate.

Each result has its own input pages. Positions across the two studies do not imply matched cases.

Input 1

English input: I deleted the stale branch deliberately.
Ainglish input: I deleted the stale branch on-purpose.

Input 2

English input: Mara skipped the flaky test deliberately.
Ainglish input: Mara skipped the flaky test on-purpose.

Input 3

English input: The operator disabled the noisy alarm deliberately.
Ainglish input: The operator disabled the noisy alarm on-purpose.

Input 4

English input: The agent selected the slower fallback deliberately.
Ainglish input: The agent selected the slower fallback on-purpose.

Input 5

English input: We delayed the release deliberately.
Ainglish input: We delayed the release on-purpose.

Input 6

English input: Nico revoked the expired token deliberately.
Ainglish input: Nico revoked the expired token on-purpose.

Recorded input digest: da8ed3dd1b478c76b590b2a454a574676c22593a78821f45c184c1779a4df1d1

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.

Declared population, method and retained outcomes

Compared with: Ainglish adverbial pin versus lossless round-trip gloss
Tested population: 24 frozen disjoint on-purpose/by-accident pairs, Saturnia replication
Unit tested: complete message
How results combine: equal item mean, then maximum tokenizer mean

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

Not recorded

Reader or tokenizer results

[
    {
        "model": "cl100k_base",
        "value": -1.5
    },
    {
        "model": "o200k_base",
        "value": -1.5
    },
    {
        "model": "p50k_base",
        "value": -1
    }
]

Condition results

Not recorded

Exact result and immutable specification Experiment history

Attempt e6546a80-20fc-4d14-9d75-cf87ee7574bf
Content 624e09d460abb232882eaa5befe910f2486230a8058bb87a7be75cea4de21bd3

Different wording, readers, exposure or populations can legitimately produce different results. A visible reference is not training the model’s weights. Current models and tokenizers have learned English; future Ainglish-trained performance remains a research question.