# Suite32 SELECT: scientific specification

## Objective

Establish whether **historically verified target-conditioned allocation of labelled training observations** improves faculty-specific early course-failure warning, while retaining valid faculty-native predictors that are removed by the all-faculty common-column intersection. This is a *data-policy hypothesis*, not another Council-anchored correction model.

## Fixed task

One semester-origin course registration: `student ID × course ID × target_NHHK`; binary `grade10 < 5`. The baseline historical caches have already performed earlier cross-fitting and feature engineering. This package cannot retroactively prove the origin availability of all upstream fields. Target labels are used only after prediction for scoring. Inference uses one fitted HGB; no other model predictions are inputs.

## Paired causal contrast (observational learning-policy experiment)

Hold fixed the target registration rows, target labels, semester cutoff, HGB family, feature contract and donor mass. Compare target only against target plus donor `s` for that target origin. Per-origin directed AP delta is diagnostic and may be noisy. Historical data availability, failed masks and grading/curriculum semantics remain potential confounders; an observed AP delta is **not** evidence of causal similarity between faculties.

Donor policy `a_{ts} >=0`, sum=1 over admitted sources, with total mass `rho` and zero-borrowing option:

`L_t(theta) = (1-rho)*mean_{i in target} CE_i + rho*sum_s a_ts*mean_{i in source_s} CE_i`.

Rows are capped per donor and normalized to preserve source mass. In `pool_all`, mass is proportional to eligible pre-cap source rows, while in `pool_equal` it is equal over donors. Donor sampling seeds are stable across model variants.

## Chronological selection and next-origin replay

Let each historical origin be an old semester in the training cache. An origin is eligible only with minimum earlier local training support and target validation positives/negatives. For each target and earlier historical origin, calculate per-donor validation AP uplift using only earlier training labels. At origin `u_j` a replay policy screens donors using only origins `u_1,...,u_{j-1}` and evaluates combined policy at `u_j`; outcome at `u_j` cannot inform selection for that same `u_j`. The final target-period model can use *all* eligible historical origin evidence for source choice because every origin precedes the outer target. Admission requires conservative replay validation. If no evidence suffices, local-only HGB is used.

This evaluates admission of a **selection procedure**, not the retrospective performance of an oracle choosing a source after seeing outer-fold labels. The default replay minimum is one independent later origin; two or more are preferable where support allows.

## Information experiments

- Local common (shared-field intersection + 17 course-target features).
- Local native (retains all cached target-native feature names/values).
- Selective common.
- Selective native (one model; donor native-only positions `NaN`).

The contrast `local_native-local_common` isolates the value of native fields under a fixed learner; `selective_common-local_common` isolates learned borrowing under common features; `selective_native-local_native` isolates donor utility with native inputs. Difference-in-differences between these arms describes interaction between information and borrowing, but should not be interpreted causally without matched sample/protocol controls.

Native columns with missing donor values are not filled with artificial zero. HGB treats these as genuinely **unavailable in that donor representation**; a future method may explicitly model applicability. Source missingness is a separate issue: a cache NaN mask cannot retroactively restore statuses lost upstream when zero-imputed. When source evidence becomes available, rerun the same matched arms with genuine masks.

## Required negative controls and stopping rules

Control whole-population borrowing (`hgb_pool_all`), equal-faculty donor mass (`hgb_pool_equal`), prespecified lexical language donor prior (`hgb_language_prior`), seeded random donor (`hgb_random_donor`), native local and common local. Report all—not just whichever looks best.

Stop novel faculty grouping if directed edges vanish across origins, source-size controls or later target periods; do not infer compatibility from clustering raw AP values. Stop observation-mask claims if true statuses cannot be reconstructed. Stop expanded architectures if the strongest native local HGB and selective-source controls do not establish a reproducible signal. Negative donor policies must route to local-only.

## Metrics and audit boundaries

Primary: unweighted faculty-fold AP with foldwise and facultywise comparison; separate 13-faculty Council support. Secondary: AUC, F2 at specified alert budgets, F1, Brier, ECE, dev-frozen threshold operational reports, TP/FP/FN/TN. Report all source supports, calendar origins, positive coverage, source mass, ESS, positive ESS, donor selection stability and rejected donors. Council and historic Suite25 results are marked **metric-only** because older prediction arrays are not bundled; observed numerical differences against them are not student-paired experiments.

## No architecture novelty claim

Directed source screening and weighting have related prior work in transfer learning and domain weighting. The distinctive AQ contribution, if measured, would be stable directed cross-faculty information borrowing under chronological testing with an independently verified observation and native-schema contract. The code itself does not prove that claim.
