# Suite26 frozen design

## Research question

Can a single framework combine the inductive bias that repeatedly wins on AQ's engineered current-state table with the student's own labelled academic trajectory, without depending on Committee as an input?

Suite26 specifically tests whether the two strongest surviving mechanisms from earlier experiments are complementary **inside the representation**:

1. tree-native current academic risk state;
2. student-specific historical evidence and observed outcomes.

It then tests a third, more novel hypothesis: historical performance can be summarized as a multidimensional student capability state and compared against a learned target-course requirement state.

## Architecture

```text
                        CURRENT STUDENT–COURSE QUERY
                        136 common + 17 course features
                                      │
                     ┌────────────────┴─────────────────┐
                     │                                  │
                     ▼                                  ▼
             SINGLE EXTRATREES                    raw projection
          score + compact leaf IDs                       │
                     │                                  │
                     └──────────────┬───────────────────┘
                                    ▼
                         TREE-NATIVE CURRENT STATE
                                    │
                                    ├──────── target course representation
                                    │                       │
                                    │                       ▼
                                    │             requirement r_c / difficulty d_c
                                    │
                                    ▼
                     SAME-STUDENT EPISODIC MEMORY
            only events with event_semester < query_semester
               course_i + tree state_i + grade_i + fail_i
                     + recency_i + same-course indicator
                                    │
                         target-conditioned attention
                                    │
                  ┌─────────────────┴─────────────────┐
                  ▼                                   ▼
             memory state m_s                 latent ability θ_s
                  │                                   │
                  │                                   ▼
                  │                           θ_s versus r_c
                  │                             ability logit
                  └─────────────────┬─────────────────┘
                                    ▼
                           NON-ADDITIVE INTERACTION
            [current, memory, course, θ, r, current×memory,
             |current-memory|, θ×r, ability_logit, support]
                                    │
                                    ▼
                          bounded joint correction
                                    │
                  internal single-tree logit + correction
                                    │
                                    ▼
                              P(grade < 5)
```

## Why a single internal tree instead of Committee

Committee is intentionally excluded from the primary architecture. A single ExtraTrees model supplies the tree-native current-state partition and current risk score. This keeps Suite26 a unified framework and lets us answer whether student memory can make one tree-native system competitive with the older multi-model Committee/Council family.

The existing Committee is loaded only for reference comparison and one post-hoc blend ablation.

## Tree leakage / overfit controls

The tree is a supervised internal encoder, so Suite26 explicitly limits same-row target leakage:

- selection tree is fit on selection-training rows only;
- validation rows are unseen by the selection tree;
- target rows are never used to fit any tree;
- for rows used to fit a forest, the neural model receives **OOB tree probabilities** rather than in-sample forest probabilities;
- leaf embeddings receive leaf identity plus support/depth only, not the leaf's training failure rate;
- temporal holdout determines epochs exactly as in recent suites.

The leaf partition itself is still supervised; this is intentional because it is the tree-native representation being tested. It is part of one sequential composite estimator, not a target score imported from the final holdout.

## Student episodic memory

Only the same student's previous course outcomes are primary memory candidates. Suite25 showed this source overwhelmingly dominated generic state, same-course and structural candidate pools.

For every query `(student s, target course c, semester t)`:

```text
H(s,t) = all rows for student s with semester < t
```

If there are more than 32 events, Suite26 keeps 24 most recent events and preserves older failures before filling remaining slots. Same-semester rows are never visible.

Attention keys deliberately exclude outcome labels. Relevance is inferred from current/tree state, historical tree state, historical/target course representations, recency and same-course identity. Historical labels and grades enter the **values**, which prevents relevance selection from trivially keying directly on the outcome.

## Latent ability–requirement interaction

Each attended historical event emits a `K=16` ability evidence vector. Attention pooling gives

```text
θ_s ∈ R^16
```

The target course encoder produces

```text
r_c ∈ R^16          course requirement
 a_c ∈ R^16_+        course discrimination
 d_c ∈ R             course difficulty
```

and

```text
z_ability = a_cᵀ(θ_s - r_c)/sqrt(K) - d_c
```

This is IRT-inspired but not classical IRT. `θ_s` is query-conditioned and learned from the student's actual transcript; `r_c` is learned from course descriptors. The ability logit is both an input to the final interaction and a weak auxiliary supervised head.

## Staged variants

### A — `suite26_tree`
Single ExtraTrees current-state baseline.

Tests whether the internal tree base itself is competitive enough to support a unified framework.

### B — `suite26_tree_memory`
A + target-conditioned episodic memory.

Tests the strongest repeated premise from Suites22–25: the student's own historical experiences and outcomes add information beyond current state.

### C — `suite26_tair` — PRIMARY
B + latent ability–requirement interaction.

Tests whether a learned capability-vs-course-requirement mechanism improves direct episodic memory rather than explicit prerequisite/graph structure.

### Diagnostic — `suite26_joint_head`
The full C representation also trains a fully unanchored head. This is not another expensive model. It lets us determine whether preserving the tree score is helping or suppressing the new joint representation.

### Secondary ablation — `suite26_committee_blend`
Chronology-selected scalar blend between Committee and `suite26_tair`. It is never the primary scientific result.

## Frozen training

- faculty-balanced sampling, natural within-faculty class prevalence;
- AdamW `lr=8e-4`, weight decay `1e-4`;
- batch 192;
- at most 12 epochs, min 3, patience 3;
- BCE + 0.12 within-faculty pair ranking;
- 0.20 auxiliary joint-head BCE;
- 0.05 auxiliary ability-head BCE;
- 0.001 bounded-correction L2 penalty;
- latest three historical blocks as validation;
- refit selected epoch count on all permissible history;
- Platt calibration fitted only on the calibration portion of historical validation;
- blend authority chosen only on the separate policy portion of historical validation.

## Frozen diagnostics / falsification

Suite26 must print and save:

- A→B and B→C AP/F2 deltas by faculty/fold;
- no-memory, no-label, no-grade, no-recency, no-tree-leaf, no-ability, uniform-attention and exact-tree ablations;
- same-student history shuffle within faculty+query-semester (chronology preserved);
- weighted historical failure/grade evidence for actual failures vs normals;
- attention mass on prior failures/passes, same/other course and grade bands;
- ability-logit distribution for target failures vs normals;
- tree correction distribution for failures vs normals;
- history-support strata and support-vs-gain correlation;
- history context size sensitivity `k={4,8,16,32}`;
- cold/new student/course regimes;
- tree feature importances;
- Committee complementarity and optional blend;
- operational fixed/drift thresholds, confusion matrices and student-cluster bootstrap CIs;
- all earlier Suite14/16/17/18/20/21/22/23/24/25 references plus Council/V2 wherever available.

## Predeclared interpretation

Strong support requires all of the following patterns, not merely one good macro number:

- B materially improves A across a broad majority of faculty-fold cells;
- shuffling or removing the student's historical evidence materially hurts;
- C improves B if latent ability/requirement is genuinely useful;
- primary standalone result approaches or exceeds Committee and materially narrows the Council gap;
- no large faculty regressions are hidden by macro averaging;
- operational F2/recall do not collapse;
- the unanchored head indicates whether the tree-preserving skip is useful rather than assumed.

If B works but C does not, keep tree + episodic memory and reject the latent-trait premise. If neither B nor C materially improves A, then our strongest relational signal cannot be exploited by this tree-native joint representation and the direction should be reconsidered rather than repeatedly tuned.
