0082 — scipy joins the allowed permissive references, for the Kolmogorov branch table

Status: accepted · Date: 2026-09-05

Context

Decision 0003 allows "permissively licensed implementations as a behavior reference only," naming rapidfuzz (MIT), jellyfish (MIT), textdistance (MIT) and scikit-learn (BSD-3). It does not name scipy — no package before Lodestar.Stats (#442) had reason to consult it.

KolmogorovSmirnov.TwoSample needs the finite-sample Kolmogorov distribution's survival function, P(D_n > d). Durbin's (1968) recursion and the Marsaglia-Tsang-Wang (2003) scaling give the exact route; Pelz-Good's asymptotic expansion gives the other end. Between them is a dispatch problem the published papers do not settle: which of several exact methods, direct formulas and asymptotic approximations applies for a given (n, d), and where each cedes to the next. Durbin's paper does not say when its own recursion stops being the cheapest correct answer against a direct combinatorial route; Pelz-Good's does not say at what n·d² an asymptotic expansion becomes accurate enough to prefer over an exact but slower one. Answering that from first principles for every one of the ten rows Internal/Kolmogorov.cs's dispatch table carries (n > 140, n·d² < 2.2 vs ≥ 2.2 vs ≥ 370, n·d^1.5 ≤ 1.4, and so on) would mean deriving error bounds for four numerical methods against each other — a research project, not an implementation task.

scipy's own _ksstats.py (scipy.stats, BSD-3) already carries exactly that dispatch, _kolmogn, tuned against decades of use. Task 8 (.superpowers/sdd/2026-09-05_0442_lodestar-stats/task-8-report.md, finding 4) found and corrected a case where this had been done without saying so plainly: the comment on DurbinCdf's scaling claimed no reference implementation was transcribed, when the scaling constant (ScaleBits = 128, scipy's own _E128/_EP128, _ksstats.py:73-75) and the rescale trigger's placement did follow _ksstats.py structurally, not merely arrive at the same answer independently. The mathematics — Durbin's recursion, Marsaglia-Tsang-Wang's scaling identity, Pelz-Good's expansion — is independently implemented from the published papers; the decision of which branch runs where, and the specific thresholds bounding each one, is read from scipy's dispatch table, attributed as such in the file's own long comment.

Decision

scipy (BSD-3) is added to decision 0003's named list of allowed permissive references, alongside scikit-learn — the same licence, the same "behavioural reference only" rule, and the same "we reproduce inputs/outputs and analogous naming, never the source" boundary. This is recorded as its own decision rather than by editing 0003's body: docs/decisions/README.md states the convention this repository already follows — "a decision record is not edited; an amendment is its own record" — and tools/check_adr_immutable.py enforces it mechanically for every ADR that already existed before this pull request. 0003 stays exactly as accepted; this decision is what a reader follows to see that its rule was extended and why.

What was read from scipy, precisely. Internal/Kolmogorov.cs's branch conditions and their order — the ten-row table in the file's own comment, and the two scaling constants inside DurbinCdf — are scipy's _kolmogn's dispatch, read and reproduced structurally. No scipy source line is transcribed: the C# is an independent recursion (DurbinCdf) collapsing scipy's first three closed-form branches into one general one, verified by derivation rather than copied algebra (the file's own comment shows the derivation for n·d ∈ (0.5, 1], cross-checked against scipy numerically, n=2, d=0.49 → 0.5392 on both sides). What crosses over is which formula applies where, which is behaviour, not code — exactly what 0003 already permits from a permissively licensed reference, extended here to name the specific package doing the permitting.

Consequences

• docs/decisions/README.md's row for 0003 gains a pointer to this decision in its relationships column — one-sided, not a cross-reference: 0003 was already accepted when this was written, so tools/check_adr_immutable.py refuses a diff that edits it, and 0003's own body says nothing about scipy. The index carries the back-reference instead, the same shape 0057 and 0058 already use for an amended decision that cannot name its own amendment.
• A future package that reads scipy's own dispatch logic to settle a branch-selection question — rather than a formula it derives independently — cites this decision rather than reopening 0003 or writing its own.
• Internal/Kolmogorov.cs's long comment remains the concrete, line-numbered record of what was read (_ksstats.py:73-75 for the two scaling constants, the ten-row table for the dispatch); this decision is the policy that comment operates under, not a restatement of it.