# Suite15 — Conditional Information Lattice Audit

Suite15 is deliberately **not** another MoE candidate. It answers the prerequisite question for the next framework:

> Given a strong chronology-safe E0, which raw structured evidence sources still contain incremental predictive information?

## Primary experiment

For `V1=student trajectory`, `V2=course/curriculum`, `V3=student-course relational history`, and `V4=fine-grained mastery`, Suite15 trains identity-initialized residual probes around a frozen E0 and evaluates all `2^4 = 16` subsets. There is no gate, routing, authority prediction, evidence exchange, or expert competition.

Every probe receives the same shared controls (faculty + history/support depth) and the same anchor logit. Raw structured tensors are used before the Suite12–14 neural expert encoders. Relation/component categories are represented by exact binary codes plus explicit masks to preserve categorical identity without multi-gigabyte one-hot matrices.

## Leakage protection

The audit-training E0 score is **not in-sample**. For each temporal block, E0 is selected/refit using strictly earlier blocks and then predicts the next block. The earliest rows without sufficient prior blocks are excluded from conditional-probe training. The target E0 is selected on historical data and refit on all permitted history.

## Secondary diagnostics (folds 6–7 by default)

- Committee-anchor single views + all views, for the empirical E0 absorption index.
- Full-context versions of V1–V4, to detect whether gains come from old context rather than raw structured evidence.
- Matched HistGradientBoosting control/view probes, to check that an MLP-negative result is not merely an MLP inductive-bias failure.
- Student-cluster bootstrap confidence intervals by `IDSV`.

## Key outputs

- `information_lattice.csv` — macro holdout performance for the full E0 lattice.
- `faculty_fold_metrics.csv` — AP/AUC/F1/F2/F2@20 for every faculty/fold/configuration.
- `view_shapley.csv` — exact 4-view Shapley AP attribution.
- `pairwise_synergy.csv` — predictive pair interaction.
- `absorption.csv` — Committee gain versus E0 gain.
- `bootstrap_ci.csv` — student-cluster AP-difference confidence intervals.
- `tree_probe.csv` — matched tree-family confirmation.
- `correction_diagnostics.csv` — correction magnitude and hardest-E0-row BCE improvement.
- `holdout_gain_drift.csv` — fold6→fold7 conditional gain drift.
- `information_verdict.json` and `next_experiment_hints.json` — compact planning summaries.

Per-fold score files are stored as `fold_N/target_scores.csv.gz`. E0 anchors and probe fits are cached so an interrupted run can resume.

## Default server paths

```text
Canonical Suite5 cache:
/root/disk_3tb/earlywarning/30072026/results/cross_faculty_residual_moe/cache

Structured schema-v4 evidence cache:
/root/disk_3tb/earlywarning/30072026/results/suite12_structured_cache

Suite15 results:
/root/disk_3tb/earlywarning/30072026/results/suite15_conditional_information_audit

Log:
/root/disk_3tb/earlywarning/30072026/log/suite15.log
```

## Run

```bash
cd /root/disk_3tb/earlywarning/30072026/suite15
bash run_suite15.sh
```

The default run executes the 16-state E0 lattice on folds 1–7 and the expensive secondary diagnostics on folds 6–7. Use `--no-tree_probe` or `--bootstrap_reps 0` only for development/debugging runs, not the final experiment.
