0004 — What is written here and what is delegated

Status: accepted · Date: 2026-09-20

Context

This project's thesis is a claim about .NET, and a claim about .NET has to be checked rather than assumed. Nineteen records answered it one capability at a time — the tokenizer, the sparse decomposition, the explained variance, the GLM, the forecast, the clustering algorithms, the splitters, the kNN kernel — and a reader who wanted the axis read nineteen records and assembled it. This record is the axis, with one row per capability decided.

The method the nineteen converged on, in the order it is applied:

1. Check this repository first. A roadmap once proposed a metrics phase that had already shipped.
2. Read the incumbent's exported surface, never its README. A MetadataLoadContext over the package's own assemblies, today tools/survey.cs, which names its counting basis. A search of NuGet alone closed two lots that a reading then reopened, and reopened two the search had left open.
3. Read the licence from the package, not from the repository, and record an empty search rather than omitting it.
4. Name the target frameworks. A library that installs only on net8.0 answers nothing for the netstandard2.0 half of what ships here.
5. Date the reading. Every verdict below was taken between 2026-08-30 and 2026-09-16, and each is a statement about what was published on its day.

Three of the nineteen corrected an earlier one, which is why the method has the shape it has. 0074 replaced 0059's NuGet searches with a reading of what the packages export. 0129 withdrew three absence claims this repository was publishing, amending 0105 and 0116. 0140 measured 0060's question on a named machine and reversed its direction. What follows states the corrected truth, not the first one.

Decision

Native code is written only where .NET has no maintained equivalent at the reference's parity, and no Python runs at runtime. Measured package by package, that gap is consistently the apparatus around a computation rather than the computation. The tokenizer loader and not its encoder, the regression's inference table and not its coefficients, the time-series diagnostics and not the forecast, sparse decomposition and not dense.

A capability is decided by three questions, in this order.

Is there a free, permissively licensed, maintained .NET incumbent at the reference's parity, on every framework this repository ships? All five clauses bind. Accord's GLM is complete and LGPL-2.1. Numerics.NET carries the time-series diagnostics and is commercial. NumFlat's PCA and Gaussian mixture are MIT and install on net8.0 alone. cs-glm has a real stack and cannot be consumed as published. Cortex.TimeSeries' ARIMA returns a pure autoregression's coefficients under an ARIMA name. Each is a different reason, and none of them is a judgement of quality.

Is there a parity target that pins its own answer at the tolerance a corpus is compared at? The corpora agree at 1e-9. A reference whose own optimisers disagree with each other by more than that cannot be frozen: what a corpus would record is one optimiser's stopping point, and a correct implementation reaching a better optimum would fail it. This is what separates the multinomial logit from the ordered model, and PanelOLS from MixedLM, with no appeal to how useful either is.

Is there a caller? A void that is real, provable and unclaimed is recorded and left. Publishing later is always available; unpublishing is not.

capability	verdict	why	from
tokenizer.json and spiece.model loading	written here	no .NET tokenizer reads the artefact a user has: every Microsoft.ML.Tokenizers 2.0.0 factory takes a vocabulary file, a spiece.model or constructed objects	0068
the encode kernels behind that loader	written here, and behind the incumbent	ordinary code losing to an available alternative is a defect to fix, not a reason to delegate the loader with it	0068
sparse truncated SVD and NMF over a CsrMatrix	written here	Math.NET's QR and SVD are dense and its sparse-SVD request has been open since 2013; parity is provable because Ω is an input rather than a seed	0059, 0072
a portable seed for that draw	not written	reproducing MT19937 buys parity the explicit Ω already gives, and a wrong Mersenne Twister fails quietly	0072
dense PCA projection	delegated to Microsoft.ML on every target, NumFlat on net8.0 and above	ML.NET projects from a netstandard2.0 package and consumes a sparse vector in O(nnz); centring a CsrMatrix densifies, so this package refuses PCA over sparse input by name	0059, 0116
the variance a component explains	written here	ML.NET's PCA exports fourteen members and no eigenvalue; the free netstandard2.0 incumbent that does report it, Meta.Numerics under MS-PL, was found after the type had shipped against a frozen scikit-learn corpus it has never been held to	0116, as amended by 0129
the ten hypothesis-test families	written here	Meta.Numerics carries eight of them on netstandard2.0 under MS-PL, so the gap is not existence: it is agreement with scipy at 1e-9 through a frozen corpus, and a span API under Apache-2.0	0129
the generalized linear model and its table	written here	Accord ships the whole thing under LGPL-2.1 and archived; ML.NET is binary logistic only, with no link function as a concept and its standard errors behind a native MKL dependency; cs-glm installs nothing	0104
HC0–HC3 robust covariances	written here, before the rest of the tail	the package sells the inference table, and a standard error computed under an assumption nobody checked is wrong and silent, where a missing family stops a fit visibly	0115
WLS, GLS, negative binomial, Gamma	written here, in that order, each opening when the one before ships	they are the same table under a different variance assumption; inverse Gaussian and Tweedie wait for a caller	0115
the multinomial logit	written here	Accord's is the one incumbent and is LGPL-2.1; statsmodels' analytic Newton reproduces itself to 2.1e-15, so the corpus holds	0136
the ordered model	not written, void recorded	nothing in .NET fits one, and the reference defines neither score nor Hessian: its default stops 2e-4 short and even Newton moves 1.6e-8 when refitted from its own optimum	0136
regularised fits — lasso, ridge, elastic net	delegated to Microsoft.ML's LBFGS trainers, with the penalty scaled by the row count	first-party, MIT, and it reaches the same optimum and the same selected variables; the reference publishes no inference table for a penalised fit, deliberately, so there is no apparatus left to add	0137
mixed and hierarchical models	not written, void recorded	no .NET package fits one, free or commercial, and the reference's five solvers disagree by 1e-5 on a coefficient and 3e-3 on the likelihood when a variance sits on its boundary	0130
time-series forecasting, seasonality and anomaly detection	delegated to Microsoft.ML.TimeSeries	first-party, MIT, maintained, 2.85 M downloads; writing an SSA forecaster beside it is the rewrite this project exists not to do	0105
the time-series diagnostics — ACF, PACF, Ljung-Box, ADF, KPSS, seasonal decomposition	written here, in Lodestar.Stats.TimeSeries	free and permissive, a partial autocorrelation, a stationarity test and a seasonal decomposition exist only in the unadopted Cortex.TimeSeries; they are their own package because ADF's lag search is an OLS, which Lodestar.Stats must not drag in	0105 as amended by 0129, 0133
ARIMA, SARIMAX, state-space estimation	not written, void recorded	no free incumbent estimates one — Cortex.TimeSeries' ARIMA returns the pure autoregression's coefficients — and the reference's own optimisers part by 2.6e-4, one of them reporting non-convergence	0134
vector autoregression	not written, and the one of the four that could be	VAR(...).fit(p) is equation-by-equation least squares on stacked lags and matches numpy.linalg.lstsq at 0.0; no .NET package estimates one	0134
instrumental variables — 2SLS, LIML, two-step GMM	not written; no incumbent, writable at parity	seven .NET libraries, three commercial, export nothing that fits one; linearmodels 7.0 reproduces each closed form at 1e-15. Iterated GMM is out: it moves with its tolerance at 6.6e-6	0135
panel estimators — PanelOLS, BetweenOLS, FirstDifferenceOLS, RandomEffects	not written; no incumbent, writable at parity	the same reading; the small-sample factors are linearmodels' own and a corpus has to pin them rather than assume statsmodels'	0135
k-means	written here	ahead of both free incumbents on the machine below, and the only other free netstandard2.0 k-means takes no seed and no starting centres	0131
DBSCAN	written here	the one free .NET DBSCAN is planar: its Point carries X and Y and nothing else	0131
agglomerative clustering	written here, with Aglomera as its measured incumbent	Aglomera covers every linkage on netstandard2.0 and has had no release since 2020 — the "no maintained equivalent" this project writes against — and has never been compared with scikit-learn	0131
HDBSCAN	delegated to HdbscanSharp	MIT, netstandard2.0, released 2025	0131
Gaussian mixture and k-medoids	delegated to NumFlat on net8.0; the mixture is a scoped gap below it	NumFlat covers both where most callers are, and k-medoids is not in scikit-learn at all	0131
MiniBatchKMeans	not written	its batches are drawn inside the fit and no parameter hands them in, so no run is reproducible from .NET	0131
spectral clustering	not written until an eigensolver exists	it needs the leading eigenvectors of an affinity Laplacian, which is not the small dense Gram matrix the explained variance solves	0131
KFold, StratifiedKFold, train_test_split	written here, first of the preprocessing lots	no .NET splitter reproduces scikit-learn's folds, the maintained first-party one cannot stratify at all, and Lodestar.Conformal's guarantee is a property of how the split was made	0132
scalers, encoders and the imputer	written here	ML.NET has all of them and welds each to an IDataView and a column name; what is added is the same answer on an array, at the reference's category order and unknown-value handling	0132
SMOTE and imbalanced resampling	not written, void recorded	the emptiest square on the board and the least provable: the one NuGet hit wraps a framework abandoned in 2019, and no parameter hands in the draws	0132
the kNN dot kernel	written here	ahead of TensorPrimitives by 3–7× on one named machine and behind it by at most 23% on another, and the dot is 23–49% of a query on the two; delegation is off the table until a machine shows it ahead by more than this one shows it behind	0060 as amended by 0140
MinHash, LSH and SimHash	not written	the sketch, the banded index and the threshold-to-(b, r) solve all ship, in MinHashSharp, SimhashLib and LuceneSharp's MinHashFilter	0059, 0074
BM25 over an in-memory CsrMatrix	written here	the Block-Max WAND top-k the lot was proposed for ships in LuceneSharp.Core; what does not ship is BM25 with no Directory, no IndexWriter, no codec and no NLP dependency	0074
keyword extraction — RAKE, TextRank, MMR	written here	a C# YAKE exists and is not written; a search for rake textrank returns zero packages	0059, 0074
BK-tree	written here; VP-tree kNN delegated to vptree	vptree is generic over any metric and takes our distance kernels, and a search for bktree returns nothing that contains one	0074

Where each measurement holds

A ratio belongs to the machine and the window it was taken in, and to nothing else.

• The kNN kernel, twice, in opposite directions. On a hosted runner — 4 cores, AVX2, .NET 10.0.11, three rounds of nine interleaved runs — TensorPrimitives was 1.09–1.23× ahead on the dot and the dot was 49% of a query. On an AMD Ryzen 7 8700G (AVX-512), Ubuntu 26.04.1, .NET 10.0.12, pinned to one core, five runs of nine, our kernel is 3–7× ahead (dot ratio 0.14 with the 512-bit path on, 0.32 with it off) and the dot is 23–33% of a query. Neither machine is declared the true one; both readings are why the kernel stays.
• Clustering, one machine. AMD Ryzen 7 8700G, default job: Lloyd's iterations from the same centres 1.52–3.66× ahead of NumFlat's, and a whole default fit 5.5–13× cheaper than NumFlat's and 4.1–20× cheaper than Meta.Numerics'. Every pair agreed on the centres before it was timed.
• The tokenizers, no machine. Microsoft.ML.Tokenizers encoded the same artefacts in 54.33 ms against our 112.36 for WordPiece and 56.94 against 682.64 for SentencePiece, both sides returning identical ids. Those are container timings and carry no machine, so the ratios are not published as speeds; the allocation beside them — 3.55 MB against 118.84, 3.09 MB against 519.51 — is a property of the code path and does not need one.
• Agreement, which needs no machine. Over the same Ω, a step-by-step reimplementation of randomized_svd reproduces U, s and Vᵀ at exactly 0.0. ML.NET's LBFGS trainers reach statsmodels' regularised optimum with the same zeros, 1.8e-5 apart because their weights are float. The two-sided normal p-value is ChiSquaredSf(z², 1) exactly, agreeing with scipy at 1.8e-16 at z = 0.5 and 1.1e-13 at z = 37.

Consequences

• A delegation is a promise to keep reading. Microsoft.ML.TimeSeries has a 6.0.0 preview out; if it grows an ACF, the diagnostics' scope shrinks and this record is superseded rather than edited. That is the cost of delegating to something alive, and it is cheaper than duplicating it.
• An absence claim is the fragile half of every row above. Three of them were false when 0129 went looking, in documents a reader meets before any decision record — docs/migration/, README.md's incumbent table, a guide. A claim that nothing exists is corrected where it is published, not only here.
• A commercial incumbent is named, never delegated to. A migration row recommends what a reader can install without buying a licence; Numerics.NET, Dew.Stats, ILNumerics and NMath are named beside the subjects they cover so that an absence claim is not quietly true only of free software. Nothing is timed or fitted under a trial licence.
• Every not written row carries its reopening condition, and it is a caller rather than a gap: the ordered model, mixed models, SMOTE, ARIMA, VAR, IV and panel each name what the lot would be, and the two families whose reference is unpinnable also name the conformance question their caller inherits.
• A verdict of delegated moves a row in docs/equivalence.md and in docs/migration/, in the same commit as the reading. That is where a reader meets it.