Status: accepted · Date: 2026-09-20

Every algorithm here is measured against a Python reference and frozen into tests/oracles/, so parity is the default and a difference is an event. Nine records asked the same question one member at a time: this member does not answer what the reference answers — is that allowed? They answered it nine times, and a reader who wanted the rule read nine records and assembled it.

The nine cases fall into three shapes, and the shape is what decides them.

The reference is wrong in a way it does not state. jellyfish 1.2.1 diverges from the published Hamming and Jaro definitions on 62 of 1241 corpus pairs, always on combining marks, emoji or mixed scripts, and the cause is inside a Rust core nobody here can read (0005). The same library measures a Match Rating codex by its UTF-8 byte length in three places — the six-character truncation, the length-gap check and the minimum-rating lookup — so match_rating_codex("並丝七世") returns a six-character codex from a four-character word, longer than the word it truncated (0080). scikit-learn 1.9.0 raises ValueError: math domain error on mutual_info_score([], []), from an unguarded log(0) eight frames down, with no parameter guard and no docstring note (0039).

The reference's behaviour is unseeded, or it degrades in silence. sklearn.cluster.KMeans was observed breaking an exact distance tie toward the second centre in one configuration and the first in another, inside compiled Cython over a BLAS product (0093). tokenizers 0.23.1 accepts a vocabulary that declares byte_fallback without all 256 <0xXX> pieces and degrades per symbol to the unknown token — or, with no unk_token, drops the symbol and lets its neighbours merge across the hole (0063). difflib.SequenceMatcher applies an autojunk heuristic past 200 elements that is an optimisation of difflib, not a property of the Ratcliff-Obershelp metric (0006).

The reference refuses, or answers, where the answer is the caller's guarantee. MAPIE 1.5.0 raises when alpha < 1/(n + 1), and under allow_infinite_bounds=True returns a finite interval whose half-width is the largest calibration score — narrower than the level asked for, in exactly the regime where nobody would notice (0070). scipy answers (nan, nan) with a warning on a sample its own guard finds degenerate, which under nan_policy='omit' would make one input raise or not according to how it arrived (0117).

0023 is in this family for the opposite reason, and is kept in the table below because of it: BpeTokenizer.Decode used to throw on a byte sequence that is not well-formed UTF-8, where tokenizers substitutes U+FFFD. Nothing defended the strict decode — it made one-id-at-a-time decoding unusable for any text outside Latin-1 — so the code was aligned with the reference and the divergence was withdrawn. That is the axis running the other way.

A divergence from the reference is accepted only when reproducing it would mean reproducing a defect, an unseeded behaviour or a refusal the reference itself does not defend — otherwise the reference wins and the code is aligned with it.

An accidental byte/character mixup, a tie broken inside a compiled kernel, a silent degradation into a wrong token stream and a logarithm of zero are all of that kind. A rule table, a default, a parameter's presence, an exception type and a lossy decode are not: they are what a caller migrating a script depends on, and they are reproduced even where they look wrong read on their own.

Every accepted divergence carries a row in docs/equivalence.md, written in the same commit as the member. The row states what the reference does and what this does, so a reader meets the divergence in the mapping table rather than in a stack trace. Where the corpus cannot hold the case — because the reference produces no value for it — a named test pins it instead, and that test names this record.

member	what the reference does	what this does	why the divergence is accepted	record
Hamming.Distance, Jaro.Similarity, JaroWinkler.Similarity	jellyfish 1.2.1 departs from the published definition on 62 of 1241 pairs, all carrying combining marks, emoji or mixed scripts; it is neither a byte comparison nor an NFC effect	the published definition, over TextElement.Utf16Unit or TextElement.CodePoint; the oracle is generated from an explicit reference of that definition and records the jellyfish divergence count	an unspecified quirk of a Rust core, reproducible only by copying it; parity holds on every input where jellyfish computes a standard result, pinned by the classic name pairs	0005
RatcliffObershelp.Similarity	difflib.SequenceMatcher treats an element occupying more than 1% of positions as junk past 200 elements, changing ratio()	2·M/T over the recursive pairing, discarding nothing; the oracle is frozen with autojunk=False	autojunk is an optimisation of difflib, not a property of the metric; identical to the default below 200 elements and to autojunk=False at any length	0006
MutualInformation.Score	mutual_info_score([], []) raises ValueError: math domain error from log(pi.sum()), with no guard and no note	returns 0.0, as every sibling agreement metric answers on an empty input	an unguarded logarithm surfacing, not a designed refusal — it was found by a generator crashing; mutual_info_score is 0.0 on every one-class input, and empty is that shape's boundary	0039
TokenizerJsonLoader.LoadBpe, BpeVocabulary	tokenizers 0.23.1 accepts byte_fallback: true with a partial <0xXX> alphabet and degrades per symbol to <unk>, or drops the symbol outright when no unk_token is declared; it accepts any decoder shape	refuses the load, naming the first missing piece, and reads the decoder of such a file strictly — a bare ByteFallback, or exactly Sequence[Replace, ByteFallback, Fuse, Strip]	the degradation is silent and its output is a wrong token stream; refusing beats embeddings that are quietly wrong, and a complete alphabet also makes the reference's own unknown-token ordering bug unreachable rather than something to reproduce	0063
SplitConformal.Quantile at alpha < 1/(n + 1)	MAPIE 1.5.0 raises ValueError; under allow_infinite_bounds=True the regressor returns a finite interval whose half-width is the largest calibration score, and the classifier has no such flag	returns double.PositiveInfinity under either ConformalQuantileRule, carried to (-inf, +inf) by Interval and to the full label set by PredictionSet	MAPIE's finite answer under-covers silently, in the one regime where the calibration set is already too small for anyone to notice; a package promising finite-sample coverage cannot ship that as its edge case	0070
MatchRatingApproach.Codex, MatchRatingApproach.Compare	jellyfish 1.2.1 measures the codex in UTF-8 bytes for the six-character truncation, the length-gap check and the minimum-rating lookup, so a codex can grow under truncation and two 2-character codices can be refused a rating	measures string.Length throughout — equal to the character count for every codex either side produces	a byte/character mixup nothing in the 1977 description states; deriving it would mean reading it out of the reference implementation, which 0149 forbids	0080
KMeans.Fit on an exact tie	sklearn.cluster.KMeans(algorithm="lloyd") assigned the second centre in one measured configuration and the first in another; neither < nor <= reproduces both	the lowest-indexed centre, numpy.argmin's rule, implemented as a strict < over the squared distance	the choice lives in compiled Cython driven by BLAS and cannot be read, only guessed; and the tie is arbitrary in Lloyd's algorithm — a sustained one differs by a relabelling every metric here is invariant to	0093
ShapiroWilk.Test, KruskalWallis.Test under NanPolicy.Omit; ChiSquare.Contingency on a NaN cell	scipy answers (nan, nan) with a warning on a sample omission left degenerate, and chi2_contingency carries a NaN cell through	raises ArgumentException: omission is a filter and the family's own guards run afterwards, unchanged; a contingency cell is a count and its marginals are divided by	answering (nan, nan) only on the path reached through omission would make one degenerate input raise or not according to how it arrived, which is a worse contract than either answer	0117
BpeTokenizer.Decode	tokenizers decodes lossily — U+FFFD for a byte sequence that is not well-formed — and encodes strictly	the same: U+FFFD on the way out, a lone surrogate still refused on the way in	not a divergence — the strict decode was one, and it was withdrawn. It threw on 6 of 6 ids for 東京 👋 decoded one at a time, and nothing defended that. The asymmetry is the reference's own, not a policy invented here	0023

• Nine rows in docs/equivalence.md carry these cases. Two are incomplete and are corrected in this lot: jaro_winkler_similarity, which states the Winkler parameters and not the combining-mark divergence its sibling rows state, and match_rating_codex, which claims exact parity over 420 words without naming the truncation half of 0080 — the comparison row names it, the codex row does not.
• Four cases cannot be frozen, because the reference produces no value to freeze: MAPIE raises where Quantile answers an infinity, jellyfish has nothing correct past the byte-safe codex range, tokenizers accepts every file this loader refuses, and scikit-learn raises on the empty labelling. They are pinned by SplitConformalEdgeTests, by [InlineData] facts in MatchRatingApproachOracleTests, by loader tests, and by a mutual_information: null fixture in clustering_agreement.json.
• No frozen case turns on the k-means tie: the two fixtures that hinged on one were replaced with centres the arithmetic decides between.
• The oracle corpora are generated from an explicit reference of the published definition wherever the library is the thing diverging — _hamming_reference, _jaro_reference and _jaro_winkler_reference in tools/generate_oracles.py, and difflib called with autojunk=False.
• A change in a reference that closes one of these gaps does not reopen the decision by itself. The divergence was never a bet on the reference staying wrong; it was a refusal to freeze something unreadable or indefensible. What surfaces such a change is the Oracles are reproducible job, and what answers it is a new record.