# Suite35 — PACE: finalized protocol

## Goal and scope
Test whether self-supervision on historical academic events transfers to chronological course-failure ranking. Failure is grade < 5; retain all canonical enrollment rows, including no-history students. Development F1–F5; retrospective F6–F7. Faculties and folds receive equal weight after averaging seeds. Historical Suite22/25/34 comparisons are archival unless same row identities are verified.

## Encoders
All encoders share event course embeddings, grade-band embeddings, continuous grade/credit inputs and masking indicators, a 128-dimensional event MLP, and permutation-invariant within-semester mean pooling with counts. E0 DeepSets/MLP pools semesters without position; E1 two-layer GRU with elapsed-semester gap embeddings; E2 two-layer four-head causal Transformer with semester gap/position information. No trainable student IDs. PAD=0, UNK=1, MASK=2 for courses; grade bands [0], (0,4), [4,5), [5,7), [7,10]; separate unknown/mask tokens. Empty history maps to a learned vector and remains eligible. Recent 16 semesters / 32 events per semester; deterministic selection and truncation audit. Frozen random encoders and independent semester-shuffle training are controls.

## Pretraining
Grade-only, course-only, combined. Independently mask 20% of observed events per task, force at least one selected token per eligible batch. Hide all grade-derived inputs when grade masked; hide course identity, credits and attempt when course masked. Decode each masked event from its corrupted local event representation plus a causal semester representation. This is historical denoising, not future-grade prediction; same historical semester visible peers are permitted. Aggregate reconstruction by masked token, not batch. Grade/course frequency-only baselines fit on SSL training students; fixed validation masks and deterministic student split. Report losses, accuracy, per-band recall, frequency/history strata and majority baselines. Sum per-task mean masked-token CE losses with weights 1,1; optimization is raw CE, and CE/log(class count) is a separately labeled diagnostic. Train AdamW 3e-4, wd 1e-3, clip 1, 20 epochs, patience 4; pretrain checkpoint chosen by held-out reconstruction loss. A final origin refits from a fresh seed on all admissible events for the selected epoch count. No future events in vocabulary, normalization or pretraining. Curriculum changes are not inferred from numeric course IDs.

## Arms
H0 same-run full engineered-feature HGB; H1 HGB plus deterministic prefix summaries; N0 full-input MLP; N1 MLP plus supervised sequence encoder; N2 MLP plus frozen SSL encoder; N3 MLP plus fine-tuned SSL encoder; H2 HGB plus frozen SSL embeddings; HR HGB plus frozen random encoder. Optional HS HGB plus independently pretrained shuffled-semester encoder and faculty-local SSL. Run all three encoders under common SSL objectives. Architecture comparisons are information/budget matched, not claimed exactly parameter matched; print parameter counts. H1/H2/HR use exactly H0 rows. Feature normalization fits training rows only; HGB never uses its built-in random validation split.

## Chronological stages
For target origin t, choose latest earlier semester v with usable validation enrollments, observed labels before t, and earlier labeled train support. Selection pretraining uses only events available strictly before v; fit supervised models only on labels available before v. Validation inputs use only prefixes available at v. After selecting downstream epoch/HGB iteration budget using validation AP, refit from scratch on all labels available before t and use a separately pretrained/refit encoder at cutoff t. Reuse SSL fits across faculties sharing cutoff/encoder/task/seed/scope. Final training labels and input prefix availability are separately checked. Native AQ mode uses the inherited earlier-semester-grade availability assumption and global year*3+term-1 order; it requires no manual calendar. Internal ordinal timestamps do not establish actual grade publication or revision time. Portable normalized inputs still support strict availability timestamps or an explicit calendar embargo proxy. Reject inconsistent registration IDs, ambiguous attempts, invalid grades and calendar ties. Do not infer numeric grades from binary labels.

## Freeze and inference
Development runs write full configuration/data fingerprints and one protocol freeze with per-arm F1/F2 thresholds fitted from development predictions, selected architecture/task using H2 paired AP against max(H0,H1,HR), and promotion decision. Retrospective verifies immutable modeling configuration and uses frozen thresholds. No F6/F7 selection or adaptive promotion. All predeclared arms can be evaluated retrospectively, labeled diagnostic; the development winner is the sole promoted candidate. Stage all executes development then retrospective under this freeze. Warm encoder reuse across cutoff boundaries is forbidden.

## Metrics and decisions
AP, AUC, Brier, 15-bin ECE, F1/F2 (separate frozen thresholds), confusion matrix, precision/recall, 5/10/20% exact alert budgets; undefined AP/AUC have explicit reasons. Tables by faculty, fold, seed and arm; paired contrasts, seed dispersion, faculty-fold bootstrap intervals (descriptive; folds correlated), history lengths, new courses and no history. Promote only development H2 gain >= .01 against each H0/H1/HR mean and wins in >=60% of paired faculty-fold cells; require positive holdout evidence before operational recommendation. No promised improvement. Pretext success alone is not transfer success.

## Integration and outputs
The default command ./run_suite35.sh reads the established canonical joblib and Suite12 sidecar schemas directly, with expanding CSV fallback and no manual exports or mappings. Numeric grades come from target_grade10, augmented automatically by DDTBHP DiemHP/DiemHPCalc or the verified raw_index/course_history.joblib schema. All canonical labeled train rows are retained by default. Common raw columns and the 17 numeric course fields are aligned explicitly; preserved sidecars use finite CF coverage and checked permutation indices. Historical grade identities are deduplicated separately from supervised rows. Target-only identities may join earlier raw histories; current/future outcomes never enter their prefixes. Files load one cohort at a time, with byte fingerprints and a full support/alignment audit. Uploaded source establishes the contracts; actual server bytes and CUDA remain unverified locally. Single run_suite35.sh captures all stages to /root/disk_3tb/earlywarning/30072026/log; results default /root/disk_3tb/earlywarning/30072026/results/suite35_pace. Save predictions, epochs, checkpoints, artifacts, protocol freeze, reports, runtime/device/version metadata. Synthetic demo is explicitly synthetic, no research claims.
