# Suite35 implementation plan

Goal: Deliver a runnable, tested research package following DESIGN.md.
Architecture: explicit source contract; audited prefixes; shared hierarchical encoders; separate selection/refit per origin; paired downstream reports.
Tech: Python 3.10+, numpy, pandas, scikit-learn, PyTorch, joblib; unittest.

- [x] Data and chronology: normalized schema, raw mapping/export adapter, deterministic dedup, availability checks, prefix tensorization and summary control. Tests: future exclusion, string IDs, missing grades, sidecar identity, no-history retention.
- [x] Models and training: common event encoder; DeepSets/GRU/Transformer; mask leakage tests, permutation invariance and causal tests; masked pretraining, supervised scratch/frozen/fine-tune, device and checkpoint handling. Train smoke tests on synthetic data.
- [x] Metrics and reporting: defined/undefined metrics, exact alert budgets, faculty/fold aggregation, paired contrasts and development freeze, full tables and diagnostic statements. Tests: thresholds, seeds, tied alert ordering, holdout no-selection.
- [x] Runner and integration: chronological inner validation and fresh final refit, cutoff-scoped pretrained artifacts, all research arms, source audit, configuration and fingerprints, log runner, demo and resumability safeguards.
- [x] Verify complete package on CPU: invariant tests, full synthetic development/retrospective flow, all encoders/objectives/arms, checkpoint replay, corrupted-input rejection; document GPU unverified if unavailable and AQ integration unverified without real source.

Implementation is directly authorized by the user. Implement in this session; independent model and metric modules may be delegated after interface contracts are written. No external repository mutations.

## Native AQ revision implementation

- [x] Direct canonical/structured schema loading, full numeric-grade training registry, finite CF/permutation alignment and automatic expanding CSV fallback with training-fitted categorical encoding.
- [x] Automatic DDTBHP/verified raw-index reconciliation, target identity-only prior-history joins, exact grade IDs, ambiguity rejection and credit provenance.
- [x] Native semester-order assumption, no-argument real AQ CLI, immutable sources, source/config/artifact guards and automatic resume.
- [x] Unique-prefix history banks, compact row maps, minibatch expansion for trainable encoders and unique frozen/evaluation encoding.
- [x] Preserved-vs-reconstructed population-standard-deviation parity, archival provenance/comparisons and resilient single-log streaming.
- [x] Native-schema fixture training, probability-derived AP validation, complete checkpoint replay and final release verification documented separately in verification/VERIFICATION.md.
