The uncapped Laplace evidence charges a model for prior volume the data does not constrain — 29.7 nats in the case that mattered.
Nothing else in the package calls laplace_log_evidence, so flipping the default breaks nothing and removes a live footgun. It was left off because the plain expansion is the textbook one, which is a defensible reason to keep it — but it should be a deliberate, documented decision rather than an accident of which formula came first.
Migrated from the deleted docs/open-work.md (item 5).
The uncapped Laplace evidence charges a model for prior volume the data does not constrain — 29.7 nats in the case that mattered.
Nothing else in the package calls
laplace_log_evidence, so flipping the default breaks nothing and removes a live footgun. It was left off because the plain expansion is the textbook one, which is a defensible reason to keep it — but it should be a deliberate, documented decision rather than an accident of which formula came first.Migrated from the deleted
docs/open-work.md(item 5).