Module active 

Active learning — spend the human’s attestation budget wisely.

The mission frame: machine carries Data→Knowledge, but wisdom (the final judgement, the label, the ratification) stays with the human. Human attention is the scarce, expensive resource. Active learning is the theory of ranking which few items are most worth a human’s judgement — so a model improves fastest per label asked, and the person is never asked to grind through the obvious.

This is the supply-side of frugality: instead of demanding mass labelling (which burdens exactly the people this project protects), the engine surfaces the handful of genuinely-informative cases and routes them for attestation — the same RequiresHumanReview discipline as the rest of the stack.

Three classic query strategies, each over the predictions already produced by the existing estimators (crate::solvers::learning) — no new model, pure ranking:

• uncertainty — query where one model is least sure (least-confidence, margin, entropy).
• committee — query where an ensemble disagrees (vote/consensus entropy, KL).
• density — weight uncertainty by how representative a point is, so the model is not lured into labelling unrepresentative outliers.

Every entry fails closed (ActiveError); ranking reuses the engine’s information theory (crate::solvers::statistics::information). Kernel-class Reduction.

Re-exports

pub use committee::average_kl_disagreement;

pub use committee::consensus_entropy;

pub use committee::vote_entropy;

pub use density::cosine_similarity;

pub use density::information_density;

pub use uncertainty::rank_informative;

pub use uncertainty::score;

pub use uncertainty::Strategy;

Modules

committee

Query-by-committee — query the points where an ensemble disagrees. Disagreement is the signal: where independently-trained models diverge, a human label resolves the most uncertainty.

density

Information density — weight raw uncertainty by how representative a point is, so the query strategy is not lured into labelling unrepresentative outliers (which are uncertain but teach the model little about the bulk of the data).

uncertainty

Uncertainty sampling — query the points where a single model is least sure. Operates on a predicted class-probability matrix (n_samples × n_classes, each row a distribution); returns per-sample informativeness and a ranking.

Enums

ActiveError

Fail-closed errors for active-learning ranking.