MaintainerPulse

temporal prediction of dependency maintenance risk

compel damian0815/compel ● elevated

55%

no release in the next 12 months

44%

new issues will go unanswered (30 days)

94%

maintainer activity collapse within 12 months

70%

chance of a release within 12 months (survival model)

Model risk over time

calibrated P(no release next 12mo) at each historical monthly snapshot

Repository activity, last 36 months

pushes issues PyPI release

What drives this score

TreeSHAP contributions; raises risk / lowers risk

Survival curve

P(still no release) m months ahead, discrete-time hazard model

What if? — poke the model

drag a signal and the deployed models rescore this package live (implied signals move together — two years without a release also zeroes "releases, last 12mo"). The needle tracks the activity-collapse model, a logistic regression: monotone by construction, so it responds smoothly to counterfactuals where the tree models step.

maintainer activity collapse, 12mo

In plain words

bus factor: one person makes 100% of pushes

Maintained alternatives

similar packages with low predicted risk; ranked by summary similarity blended with shared-dependents overlap, validated against known migrations

package	what it is	similarity	its 12mo risk
sentencepiece	Unsupervised text tokenizer and detokenizer.	0.53	19%
transformers	Transformers: the model-definition framework for state-of-the-art machine learning models	0.52	0%
sentence-transformers	Embeddings, Retrieval, and Reranking	0.50	0%
textual	Modern Text User Interface framework	0.47	0%

Signals at the latest snapshot

days since last release	22	releases, last 12mo	3
people pushing, last 12mo	1	bus factor (top pusher share)	100%
issues opened, last 12mo	0	30-day response rate	no issues
stars accumulated	562	open vulns without fix	0

Probabilities are isotonic-calibrated on a held-out validation year and evaluated on future snapshots the models never saw (PR-AUC 0.894, precision@50 = 1.00 on the headline target).