decisions

Architecture decision records — index

Every accepted or superseded decision behind this codebase, one row per file. The status and date columns are copied from each file's own **Status:** … · **Date:** … line — correct the ADR to correct this table, not the other way around. The relationships column is a reading: what a **Status:** line does not say but an ADR's body does, either in its own words or in an > **update:** block naming what a later change made stale. An ADR's body is never rewritten to agree with a later one — see 0022 §10 for the convention a later decision uses instead.

#	Title	Status	Date	Relationships
0001	Target frameworks: net10.0 and netstandard2.0	accepted	2026-08-01 (revised 2026-08-04)	—
0002	Unicode comparison unit	accepted	2026-08-01	—
0003	Code provenance and license	accepted	2026-08-01	—
0004	Levenshtein: bit-parallel (Myers) optimization	accepted	2026-08-01 (revised 2026-08-05)	backlog partially cleared by issue #208 (not an ADR — see the two > **#208 update:** blocks)
0005	Accepted divergence from jellyfish on combining marks (Hamming, Jaro)	accepted	2026-08-01	—
0006	Ratcliff-Obershelp: difflib's autojunk heuristic	accepted	2026-08-01	—
0007	Metaphone: validation scope (real words)	accepted	2026-08-01	—
0008	Italian enza/enze: follow nltk, not the published description	accepted	2026-08-05	—
0009	The sample restores from a local feed, not nuget.org	accepted	2026-08-05	amended by 0012
0010	Stop-word lists come from Snowball, not from the nltk corpus	accepted	2026-08-05	—
0011	Persisted artifacts are versioned JSON, written with System.Text.Json	accepted	2026-08-05	—
0012	Each package versions and releases on its own	accepted	2026-08-05	amends 0009
0013	What SentencePieceTokenizer promises, and where it knowingly differs	accepted, superseded in part by 0014	2026-08-06	§1 (the oracle covers the vocabulary, not the stock pipeline) superseded by 0014; §2 (the unknown-piece penalty floor) stands
0014	Interpret the precompiled_charsmap, do not reimplement the rules	accepted	2026-08-06	supersedes 0013 §1
0015	The Sonar rules run in the build, not only after the push	accepted	2026-08-06	amended by 0019
0016	Classification metrics ship as DataNet.Metrics, not inside DataNet.Text	accepted	2026-08-06	—
0017	What BpeTokenizer promises, and where it knowingly differs	accepted	2026-08-09	—
0018	Multiclass ROC-AUC parallelism is opt-in, and the caller names the worker count	accepted	2026-08-10	—
0019	The .NET code-quality analysers run in the build too, and samples/ is analysed at all	accepted	2026-08-10	amends 0015
0020	normalize= is a projection, and ZeroDivision keeps a default per metric	accepted	2026-08-10	—
0021	Eleven regression metrics first, and multioutput as a method rather than an enum	accepted	2026-08-10	—
0022	How an added token matches, and what it costs the round trip	accepted	2026-08-10	§10 partially revised by issues #119 and #120 (not ADRs — see the > **#119 and #120 update:** block)
0023	Byte-level Decode substitutes U+FFFD instead of throwing	accepted	2026-08-14	—
0024	The weighted median averages two order statistics within scikit-learn's epsilon, not exactly at half	accepted	2026-08-14	—
0025	Quickselect, with an introselect fallback and a branchless partition, replaces a full sort for the unweighted median	accepted	2026-08-14	—
0026	R² and ExplainedVariance split their undefined cases differently	accepted	2026-08-14	—
0027	R²'s and ExplainedVariance's unweighted accumulation vectorizes only for a single output	accepted	2026-08-14	—
0028	Log1P is Kahan's identity, not Math.Log(1 + x)	accepted	2026-08-14	—
0029	BalancedAccuracy's adjusted is left to IEEE 754 at the one-class edge	accepted	2026-08-14	—
0030	CohenKappa keeps scikit-learn's expected-matrix orientation, and weighting only orders	accepted	2026-08-14	—
0031	NoSampleCorrect mirrors NumPy's float64 upcast, not requested-label accuracy	accepted	2026-08-14	—
0032	FScore substitutes tp/predicted/support algebraically, not via precision and recall	accepted	2026-08-14	—
0033	CompensatedSum is Neumaier's variant, and its SIMD lanes are not bit-identical to it	accepted	2026-08-14	—
0034	Distributional proof is admissible, and dropout is still refused	accepted	2026-08-14	—
0035	A null pre-split drives Apply as Removed with invert, not Isolated	accepted	2026-08-15	—
0036	A member may ship without an oracle, if the documentation says so	accepted	2026-08-16	—
0037	The guards run before the commit, through core.hooksPath and no framework	accepted	2026-08-18	extends 0015's reasoning to the Python guards
0038	The gate confronts an exception tag with the page that documents it	accepted	2026-08-18	—
0039	MutualInformation returns 0.0 on an empty input; scikit-learn raises	accepted	2026-08-18	—
0040	A curve is a sealed class per curve, not a record and not out-parameters; drop_intermediate's asymmetric defaults are reproduced	accepted	2026-08-18	—
0041	One sample file per public class, named after it	accepted	2026-08-19	—
0042	Phonetic encoders refuse a null word	accepted	2026-08-20	—
0043	The equality table is sized to the pattern, not to Latin-1	accepted	2026-08-20	amends 0004
0044	Compression belongs to the caller, not to the artifact format	accepted	2026-08-20	leaves 0011 untouched: the artifact on disk does not change
0045	A Console call carries its reason on the line, not in an exemption list	accepted	2026-08-21	adds a seventh guard to 0037's hook, under the rule it set
0046	check_adr_immutable.py runs in CI only, not the pre-commit hook	accepted	2026-08-20	—
0047	One bit-parallel gate per kernel, not one per alphabet	accepted	2026-08-21	amends 0043
0048	The bit-parallel gate depends on the kernel and the alphabet	accepted	2026-08-21	amends 0047
0049	Two gates per kernel, the second tested where the width is already known	accepted	2026-08-21	amends 0048
0050	The SentencePiece-BPE lineage stays a BPE model, and metaspace becomes one transform	accepted	2026-08-21	amends 0017 §3, whose byte_fallback refusal and "no path here" for Llama-2 and Mistral v0.1 both fall

What accepted means here

All fifty carry accepted. None has been rejected or withdrawn — a status this table would otherwise need a second word for. 0004 read a progress sentence (single-word and blocked shipped) where a status belongs; that sentence is now the opening line of its own ## Done section, and its status reads accepted like the other thirty-two.

Relationships not stated on a **Status:** line

Two pairs supersede or amend each other in the body only, not in the status line:

• 0009 is amended by 0012: 0012 replaced the repository-wide $(Version) 0009 bound the sample to with one property per package, in a > **Amended by 0012.** block inside 0009's own ## Decision section.
• 0015 is amended by 0019: 0019 measured false the reason 0015 gave for keeping samples/ off the analyser, in two > **Amended by 0019 (2026-08-10).** blocks inside 0015 itself, and 0019's own text says so directly — "this ADR amends 0015 accordingly."

0013's partial supersession by 0014 is the one relationship already on a status line; its body adds the detail that only §1 (the oracle's normalizer scope) is superseded — §2 (the unknown-piece penalty floor) stands, unaffected by 0014.