Files
zeekay 026ac25fcf hanzo-learn: fitting is a transformation, a model is a value, predicting is a query
A crate whose whole interface is two verbs. `Config` holds hyperparameters and has no
`predict`. `Fit` maps `Config -> Data -> Model`. `Model` is immutable, has private fields
and no public constructor, so an unfitted model is not a state to be checked for — it is
a value that cannot be spoken, and `check_is_fitted` has no job to do. `Predict::Answer`
is `f64` for a regressor and `Class` for a classifier: the two do not unify, so a
regressor and a classifier are different types rather than one type with a mode flag.
There is no `get_params`, no `BaseEstimator`, no clone protocol and no `Pipeline` — those
exist in Python because a config and a fitted model are the same object there.

Every fitted model is named by `Address`, a content address over its shape, its
hyperparameters and its learned parameters, hashing floats as IEEE-754 bits and
length-prefixing every sequence so the encoding is injective. Same construction and same
reasoning as the risk plane's published model values, so a fit can be stored, cited by a
decision, compared against a challenger and rolled back to without a registry.

Three estimators: least squares by Householder QR on centred data (QR not the normal
equations, which square the condition number; centring not a ones column, so the
intercept never enters the factorisation), binary logistic regression by Newton steps on
a strictly convex penalised objective, and gradient boosted CART regression trees.

A TREE READS FEATURES AT IEEE-754 BINARY32, and that is the algorithm rather than a
detail. scikit-learn casts X to float32 before its splitter sees it, XGBoost and LightGBM
do the same, and ai.onnx.ml's TreeEnsembleRegressor stores nodes_values as floats. A tree
never does arithmetic on a feature value — it only orders values and places thresholds
between them — so precision decides only which side of a boundary a sample falls on.
Placing the threshold at single(a)/2 + single(b)/2 instead of a/2 + b/2 moves it by 2e-8,
which is enough to send a sample down the other branch. Measured over the 40 depth-1
trees of the stumps fixture, whose roots all see the full sample set: the single-precision
midpoint reproduces scikit-learn's threshold bit-for-bit 40 times out of 40, the
double-precision midpoint 0 times out of 40.

Correctness is asserted against scikit-learn itself. Fixtures generated by 1.9.0 are
committed and read as data, so `cargo test` needs no Python and the oracle's version is a
recorded fact. 195 boosted trees / 1,977 split nodes agree node for node — split feature,
threshold and both children all exact — with leaf values to 1.3e-15 and predictions to
8.9e-16 against a 1e-10 bound; least squares agrees to 1.1e-15 against 1e-10 and logistic
to 1.2e-8 against 1e-7, with hard labels exact. Each tolerance is derived from the
mathematics and every assertion prints the deviation it observed, so a regression that
stays inside a bound is still visible. The generator refuses to record an oracle that is
not a function of its data: at scikit-learn's default min_samples_leaf a tree's structure
varies with random_state (20 structures over 40 seeds, prediction spread 3.9), so every
case pins the tie away and proves determinism over 12 seeds before writing a number. This
crate breaks split ties by lowest feature index instead, because a content address over a
model an RNG helped choose would name one draw rather than one fit.

The exhaustive split search is not lowered to hanzo-kernel, and the obstacle is specific:
a launch geometry must be stated before the launch, and across all 69 launch sites in
that crate Grid::Static is the only variant used, every geometry computed on the host from
a count it already knows. A CART node's sample count is not such a number. The histogram
formulation dissolves this — bin count is a hyperparameter, so the per-level geometry is
host-knowable — but it quantises thresholds and so cannot hold the node-for-node bar. It
is named, not built.

scikit-learn is a test oracle only. No code or docstring is taken from it, and NOTICE
records that, along with the obligation to add BSD-3 attribution if that ever changes.

The crate also carries, unchanged and unclaimed, the anomaly-detection and preprocessing
modules written concurrently on another track (isolation, neighbour, transform, scale,
impute, encode, metric, moment, split, twister, plus benches and the perf oracle). They
share lib.rs and tests/sklearn.rs with the estimator seam, so the crate cannot be
committed in halves. Their `Transform` and `Outlier` traits extend this seam without
changing `Fit` or `Predict`. One of their fixture tests,
`the_imbalanced_metrics_match_scikit_learn`, is red on a fixture key that measure.json
does not yet carry — which is why this is a branch and not main.
2026-08-04 17:03:22 -07:00
..