# Suite15 Conditional Information Lattice Audit — Design Specification

## Goal
Determine whether each raw structured academic evidence view contains predictive information beyond a strong chronology-safe E0, before building Conditional Specialist Boosting or differentiable-tree specialists.

## Primary question
For raw views V1 trajectory, V2 curriculum, V3 student-course relational history, and V4 component mastery, estimate the empirical conditional value `Perf(E0 + V_S) - Perf(E0)` for every subset `S` of the four views.

## Non-goals
Suite15 is not a new MoE, does not use gating, expert competition, authority prediction, evidence exchange, GRANDE/NODE/TEL, or sequential specialist boosting. It is a diagnostic experiment.

## Anchors
- Primary: a Committee-preserving neural E0 with the Suite14 architecture and strict walk-forward cross-fitted predictions on audit-training rows.
- Secondary: the original Committee probability mean.
- E0 is frozen during all conditional probes.
- All residual probes initialize to exact anchor identity.

## Raw views
- V1: `student_seq`, `student_mask`.
- V2: `course_target`, `course_neighbors`, `course_neighbor_mask`, `course_relation_type`.
- V3: `relation_events`, `relation_mask`, `relation_type`, plus target-course representation.
- V4: `component_events`, `component_mask`, `component_type`.
- Core views exclude old specialist-specific context. Single-view full-context probes are secondary diagnostics only.

## Shared controls
All conditional probes receive the same shared controls: faculty identity plus chronology-safe support variables (`history_semesters`, `history_rows`, `course_prior_rows`, `new_student`, `new_course`). No absolute semester ordinal is provided.

## Primary probe
For anchor logit `z0`, controls `C`, and subset `V_S`, train a compact residual MLP `delta_S = g(C,V_S,z0)` and predict `sigmoid(z0 + delta_S)`. Output layer is zero initialized; the correction is bounded. Checkpoint selection uses faculty-robust AP improvement.

## Lattice
Evaluate all 16 subsets of `{V1,V2,V3,V4}`, including the empty/control-only subset. Compute individual gains, leave-one-view-out effects, exact four-view Shapley attribution, and pairwise predictive synergy.

## Secondary probes
- Committee anchor: empty/control, V1, V2, V3, V4, and all views.
- Full-context MLP variants: V1..V4 only.
- Tree confirmation: HistGradientBoosting probes for V1..V4 and all views, always compared with a matched controls-only tree baseline.

## Temporal protocol
For each target fold, all training and model selection are historical. E0 OOF scores are generated by walk-forward fitting: a row's E0 model may use only strictly earlier temporal blocks. Target E0 is selected on the latest historical validation block then refit on all allowed history. Conditional probes use only rows with chronology-safe E0 OOF coverage.

## Metrics
Primary: mean-fold AP and faculty-level delta AP. Secondary: AUC, F1, F2, precision, recall, F2@20 for backward compatibility. F1- and F2-optimal thresholds are selected on historical validation data and frozen for target evaluation.

## Uncertainty
On holdout folds 6 and 7, use student-cluster bootstrap by `IDSV` to estimate 95% confidence intervals for delta AP. Never row-bootstrap registrations independently.

## Decision outputs
Print/save: full subset lattice, faculty/fold win matrices, Shapley AP/F1/F2, pairwise synergy, leave-one-out loss, Committee-to-E0 absorption index, tree/MLP agreement, correction magnitudes, correction by E0 confidence/error status, fold6-to-fold7 gain drift, and an automatic evidence verdict per view.

## Success interpretation
- Strong incremental view: macro delta AP >= 0.010 with broad faculty/fold support.
- Meaningful incremental view: macro delta AP >= 0.005 with broad support.
- Weak/redundant: near-zero/inconsistent gain and bootstrap interval crossing zero.
- Committee gain but no E0 gain indicates E0 absorption.
- Strong pair synergy can justify merging views in the next specialist design.
