# Suite26 — Tree-Native Encoder + Target-Conditioned Student Memory + Latent Ability–Requirement Interaction

**Primary output:** `suite26_tair`

Suite26 is a standalone AQ early-warning framework designed from the mechanisms that survived Suites20–25:

- the engineered/current-state AQ table is strongly tree-friendly;
- generic MLP / TabM / TabR current-state backbones are materially weaker than the tree family;
- the student's **own prior academic experiences and observed outcomes** are the most repeatable relational signal;
- generic graph topology, prerequisite pools, faculty modulation and exact semester-sequence backbones have not provided robust gains;
- relational information should interact with current state before the final decision rather than being only a post-hoc Committee correction.

The primary Suite26 model therefore combines a **single internal ExtraTrees current-state encoder** (not the three-model Committee), target-conditioned same-student episodic memory, and a multidimensional student-ability / target-course-requirement interaction.

## Frozen variants

1. `suite26_tree` — one ExtraTrees current-state baseline. This is **not Committee**.
2. `suite26_tree_memory` — tree-native current representation + target-conditioned same-student episodic memory.
3. `suite26_tair` — **PRIMARY**; adds latent ability–requirement interaction.
4. `suite26_joint_head` — diagnostic unanchored head from the same primary model.
5. `suite26_committee_blend` — secondary ablation only; Committee is not input to variants 1–4.

## Run

```bash
cd /root/disk_3tb/earlywarning/30072026
unzip suite26_tree_native_student_memory.zip
cd suite26_tree_native_student_memory
chmod +x run_suite26.sh
./run_suite26.sh
```

Default output:

```text
/root/disk_3tb/earlywarning/30072026/results/suite26_tree_native_student_memory
```

Default streamed log:

```text
/root/disk_3tb/earlywarning/30072026/log/suite26_tree_native_student_memory.log
```

## Important methodological details

- F1–F5 remain development; F6/F7 remain untouched holdout.
- Latest three historical temporal blocks are validation for each faculty.
- Historical memory is strictly **same student** and strictly **earlier semester**. Same-semester outcomes are forbidden because same-semester courses are concurrent/unordered.
- Selection tree is fit only on older selection-training rows. Validation is unseen by the tree and neural model.
- Production tree is refit on all permissible history only after epoch selection.
- Downstream neural training uses the forest's **out-of-bag current-state score** on rows used to fit that forest. This reduces same-row target leakage from the tree score while keeping the forest as one unified internal encoder.
- Learned leaf embeddings use leaf identities plus fixed support/depth statistics; they do **not** receive in-sample leaf failure rates.
- Target labels never enter target memory. History grades/outcomes are legitimate because their semesters are strictly earlier than the query.
- Committee is reference-only plus a secondary blend diagnostic.

## Default architecture

- current features: 136 common raw features + 17 target-course descriptors (when the canonical 14-faculty intersection is 136);
- ExtraTrees: 96 trees, max 768 leaves/tree, `min_samples_leaf=4`, `max_features=.70`, bootstrap/OOB;
- tree leaf embedding: 16-d;
- current state: 64-d;
- target course representation: 32-d;
- student episodic memory: 64-d;
- latent ability / course requirement: 16-d;
- maximum same-student history: 32 events (24 recent + older failures preserved when truncation is needed);
- primary score starts as the internal tree logit and learns a bounded interaction correction (`±2` logits) from a non-additive joint tree-memory representation;
- an unanchored joint head is trained in parallel as a diagnostic so we can test whether tree preservation helps or suppresses the new representation.

## Main result files

Per fold:

- `<faculty>_scores.csv`
- `suite26_tree_memory_selection_history.csv`
- `suite26_tair_selection_history.csv`
- `history_support_target.csv`
- `tree_feature_importance.csv`
- `memory_ablations.csv`
- `shuffled_student_history.csv`
- `memory_mechanism_diagnostics.csv`
- `attention_groups.csv`
- `history_k_sensitivity.csv`
- `history_support_strata.csv`
- `novelty_regimes.csv`
- `committee_complementarity.csv`
- `fold_summary.json`

Global:

- `all_fold_metrics.csv`
- `paired_holdout_summary.csv`
- `holdout_faculty_head_to_head.csv`
- `operational_metrics.csv`
- `paired_cluster_bootstrap_holdout.csv`
- `component_uplift_holdout.csv`
- `memory_ablations_all.csv`
- `shuffled_student_history_all.csv`
- `memory_diagnostics_all.csv`
- `attention_groups_all.csv`
- `history_support_strata_all.csv`
- `novelty_regimes_all.csv`
- `support_gain_correlations.csv`
- `history_k_sensitivity_all.csv`
- `premise_audit.csv`

Do the first real F1–F7 run unchanged. The staged A→B→C experiment is intentionally frozen so the result can distinguish a weak tree base, a failed memory mechanism, and a failed latent ability–requirement mechanism rather than confounding them with hyperparameter tuning.
