Sources, corrections and evidence gaps
The library links the official scikit-learn metric and pipeline guidance and MLflow tracking documentation. Worked counts and plots are invented teaching inputs. The retained experiment JSON is a blank template, not an executed or approved research run.
Research design and arithmetic, not model recommendations. Classification scores are not monetary returns; a registered artifact and a passing software test are not independent strategy validation.
Primary documentation
The library links the official scikit-learn metric and pipeline guidance and MLflow tracking documentation. Worked counts and plots are invented teaching inputs. The retained experiment JSON is a blank template, not an executed or approved research run.
Unresolved accountability details
A monitored publication contact, operator record and commercial disclosure were not provided. Old email strings have not been verified as working channels. Corrections require a confirmed channel; the dated source and build report preserve what was changed meanwhile.
Source functions have narrow jurisdictions
Official metric references explain mathematical definitions; chronological splitting and preprocessing references describe software behavior; experiment tracking explains storage of runs and artifacts. None supplies this publication with a genuine price history or proprietary validation result. Product documentation likewise establishes a declared mechanism rather than its success in a paid account. Our source map separates these uses so a reader can follow a link for the actual claim being discussed. A general landing page is not promoted into evidence of a precise return, latency figure or calibration grade. Mechanism review and outcome verification are separate editorial activities.
A source version must match the decision clock
For real research, retain source revision, observation identity, availability and correction policy. A date printed next to a value is insufficient if the value was published or revised later. Feature construction needs the version available before the declared decision. Label construction may use later outcomes under its frozen contract, but must not leak those outcomes into the input. Preserve missing observations and rejected joins as explicit records. A surviving row table can otherwise conceal a selective repair process. The existing chronological planner computes fixed-horizon boundaries only; it does not resolve actual data-arrival history.
Authored evidence is downloadable without pretending to be real
The default probability CSV expands all forty A rows, forty B rows and twenty C rows. Each row carries its group, constant assigned probability, binary teaching label and fictional classification. Its authored order is arbitrary, not a chronological series or independent sample. The seven compact cases share a transparent calculation contract. The filled cards identify missing fitting, evaluation and release evidence rather than attaching fake experiment hashes. Public downloads have no customer identifiers or secret payloads. Reproducing their arithmetic is useful; counting their rows as market validation would change their meaning.
Corrections need the prior assertion
When a source or calculation changes, retain what was originally stated, what replaces it and why the distinction matters. A revised source can invalidate downstream features even if a summary score is unchanged. Original article revision dates remain alongside the expansion date. Third-party page availability and terms can change after review; source review is not an ongoing monitoring promise. TODO(Darren): provide a monitored publisher contact and a lawful primary-record inventory for any future measured vendor claim. Until then, unknown account histories, instrument permissions and historical model cutoffs stay unknown rather than being manufactured for citation.