MesoMathematics / spec /navigation.md
anon-d3a4's picture
Add files using upload-large-folder tool
e4dca97 verified
|
Raw
History Blame Contribute Delete
2.95 kB

Section 4 navigation view

Status: computational object and v4.33.0 statement labels frozen.

Declaration graph

The whole-library vertex set contains non-auto/private raw declarations whose kind is theorem, definition or abbrev. A directed edge source -> target means that source consumes target: the target occurs as a constant in the source type, stored value, or both. Self references are removed and repeated occurrences are collapsed, while the type/value provenance bits are retained.

The complex-analysis test set is the induced vertex set whose defining module equals Mathlib.Analysis.Complex or begins with that dotted prefix. This is not a filter on declaration names.

Features

The six global features are log in-degree, log out-degree, log PageRank, log sampled directed betweenness, DAG layer and kernel definition height. PageRank uses damping 0.85. Betweenness uses 500 seeded source vertices with seed 42.

The seventeen internal features are the induced graph's in/out edge and neighbor counts, total and reciprocal neighbors, two-hop in/out neighborhoods, clustering, component size, type-reference in/out ratios, forward and reverse PageRank, HITS hub and authority scores, and an isolation indicator. Count and centrality transforms are fixed in navigation/features.py.

The old pilot used two lean-training-data explicit-premise ratios instead of the new native type-reference ratios. Consequently, the v4.33 result is a new pre-registered replication, not a byte-for-byte rerun of the old 23-column matrix.

Within each split, missing values are imputed by training medians and every column is standardized by its training mean and population standard deviation. No name, module, declaration kind or text feature enters a model.

Labels and validation

The external target must be a statement-level reviewed v4.33 file and must pass exact-name validation against the module slice. The two recovered v4.30 files remain historical evidence only: the old pilot used 36 positive nodes, whereas the recovered theorem-to-declaration mapping names 34 unique nodes, with a symmetric difference of 28 names. Exact-name survival in v4.33 is a necessary check, not a replacement for reviewing the corresponding statements.

Run 25 stratified 80/20 splits with seeds 20260804 + 1000 + repeat. Fit class-balanced logistic regression with an unpenalized intercept and L2 coefficient 0.1. Average precision is primary; the random reference for each split is its validation prevalence. All split indices, transformations and first-split coefficients are written into the result.

The earlier 1,549-declaration/36-positive pilot supplied the preregistered candidate names, not the analyzed graph. The current statements for all 36 names were reviewed against their labels and approved at the pinned v4.33 commit; the evidence and decision rule are recorded in labels/v4.33-statement-review.md.