0099 — Survival has no .NET incumbent, and scikit-survival is refused on its licence

Status: accepted · Date: 2026-09-10

Context

#442 called survival analysis the largest void it surveyed and held it in reserve, because largest void is not the same as best next move. #569 takes it, and two questions had to be settled before any code: what the incumbent is, and which Python library may serve as the oracle.

ADR 0074's amendment, restated by #442's closing note, requires an incumbent's exported surface to be read through a MetadataLoadContext rather than taken from its README. Decision 0096 is why that matters: reading Accord.Statistics that way replaced #566's gap claim instead of confirming it.

The reading, and what it found

There was no assembly to load. A NuGet capability search on 2026-09-09 returned:

query	packages
survival analysis	0
kaplan meier	0

The only domain in #442's survey where every search comes back empty. And the sharing claim held too: nothing in the packages shipped before this one is a survival primitive.

Where there is no assembly, the protocol is discharged by recording the searches — their queries, their date, their counts — and not by claiming a MetadataLoadContext run that had nothing to open. Per #442's own instruction, that absence is the benchmark section, and the benchmark harness's own README says so in its section 20 rather than leaving it blank.

Decision

lifelines 0.30.3 is the oracle. scikit-survival is refused.

scikit-survival is the nearest reference in any language, and its PyPI license_expression is GPL-3.0-or-later. Decision 0003's rule is unambiguous — permissive references only, never copyleft — and it is the rule that already excluded abydos and removed python-Levenshtein. The refusal is recorded here so nobody rediscovers it mid-lot, and so that the reason is on file as licence, not unavailability.

lifelines is MIT, read from the wheel rather than from the PyPI classifier. Its transitive graph was read the same way, because two of its dependencies report a licence pip cannot parse:

package	what pip reports	what the wheel carries
lifelines 0.30.3	MIT	MIT
autograd 1.9.1	(blank), License: UNKNOWN	License-Expression: MIT
autograd-gamma 0.4.2	License: UNKNOWN	MIT, in LICENSE
formulaic 1.2.2	(blank)	License-Expression: MIT, MIT classifier
interface-meta 2.0.1	MIT	MIT

That is the trap decision 0075 named when it read doublemetaphone's licence out of its own wheel: a License: UNKNOWN in legacy metadata is not a licence, and is not a refusal either.

Two costs the dependency carried, both measured

pandas drops a major version. lifelines requires pandas<3.0, and the lock pinned pandas==3.0.5, which arrived with statsmodels in #566. The resolver moves it to 2.3.3. CONTRIBUTING's rule is to regenerate the corpora and confirm they have not moved: all 127 agree, stats_ols.json — statsmodels' own output over that pandas — byte for byte included.

autograd-gamma 0.5.0 publishes no wheel. pip-compile selects it, and --only-binary :all: then refuses the whole install at every one of this repository's install sites. It is pinned to 0.4.2 in tools/requirements.txt, the newest that ships a wheel, which leaves the rest of the resolved graph unchanged.

lifelines' own Python floor is 3.11, below the 3.12 decision 0065 sets, so it does not raise ours — checked before adding it rather than after a generator failed.

Consequences

• Lodestar.Survival 0.1.0 ships KaplanMeier, NelsonAalen and LogRank, right-censored only. Left truncation, interval censoring, Cox regression and the accelerated-failure-time models are each their own lot.
• The benchmark project compares the three estimators against each other and against input shape, because there is no package to race. tools/check_bench_map.py's CLASS_DIRS gains the directory, which is what makes the nightly obliged to run it (#586).
• If a .NET survival package appears later, this record is what a reader compares it against: the searches are dated, so the claim can be re-run rather than argued about.