# Suite34 frozen experiment protocol — v1.0.0

This protocol and all executable sources are hashed before fitting. See
DESIGN_FINAL.md for the complete finalized decisions and DESIGN_ORIGINAL.md for
the supplied specification. Do not edit a running experiment or relax gates.

## Question, units and primary estimand

Does multiplicative student×course / history×course fusion improve a single
full-information failure predictor beyond additive fusion, arbitrary grouping,
and capacity-matched full-input learning? All arms use BCE, the same retained
history, preprocessing, fixed optimizer budget and target registration support.
Each faculty is fitted independently under one unified learning algorithm.

Primary: unweighted faculty–fold AP, first averaging metrics over matched seeds.
Never ensemble seed probabilities. No-positive cells have undefined AP and are
counted explicitly; a complete principal decision requires defined comparisons.
F1–F5 development; F6/F7 repeatedly inspected retrospective periods.

## Input evidence and provenance

Required old caches: cross_faculty_residual_moe/cache and suite12_structured_cache.
No newer cache or raw database export is required. Full information means the
common-raw-plus-17-course baseline contract, ordinarily 153 original channels;
faculty-native extras are outside this architecture-controlled study.

The exact CF mask, finite probability filtering and historical sidecar positions
follow the inspected Suite32 adapter. Structured target order is inherited from
that source contract when explicit target positions are absent; shape is checked
but independent raw-database target identity reconciliation is unavailable.
Original feature release-time/provenance and upstream imputation remain
uncertified. Cache nonfinite indicators do not recover original missingness.

The explicit reviewed feature catalog binds to actual cache names at preflight.
Unreviewed live fields are listed and retained in the full path only. Permute
feature bundles, never student values, for P1. Partition seed = 34000+model seed.
Categorical one-hot support is training-fitted with 32 fixed slots plus unknown;
the raw numeric code and cache-missing indicator remain in every model.

## Training

Default 128-wide, 2-block residual MLP; 32-dimensional views; rank 24; dropout .1.
A1/A2/P1 trainable and optimizer parameter counts must match. A0M is within 5%.
AdamW: lr .0003, weight_decay .001, natural-prevalence batches 1024, gradient
clip 1, maximum 18 epochs, minimum 3, patience 5. AMP optional, CPU fallback.
Seeds 29,129,229. A3 auxiliary coefficient .05, only when prespecified in CLI/config.

One outer fold's own canonical history per fit. Global semester parser derives
absolute year*3+term-1 from target_NHHK. Initial validation uses the latest three
historical global semester blocks; support-only fallback order 2,1,4,5,... .
Both fitting and validation need >=1 positive and >=1 negative. No random split.
Validation selection is equal-weight defined-semester AP; ties within 1e-6 use
smaller BCE, then earlier epoch. Minimum-epoch rule is enforced.

Fit selection preprocessing on fitting rows. Discard selection weights. Fit new
preprocessing on all permissible history, then reinitialize and refit for the
selected epoch count. Log every epoch and optimizer/AMP-skipped update count.
F7 can use F6 labels already in its historical cache under this fixed protocol.

## Warning policies

Separate faculty×arm×seed beta1/F1 and beta2/F2 policies. F1 unavailable; F2 uses
F1; ...; F5 uses F1–F4. F6 AND F7 reuse the exact F1–F5 policy. At least one
supporting failure is required. Maximize pooled earlier-development F-beta over
distinct scores; largest threshold wins ties; predict positive at p>=threshold.
No current-period or other-seed outcomes select a threshold.

Budgets 10%,15%,20%,25% use ceil(fraction*n), descending score, then ascending
canonical row_index. They are supplementary and separate from threshold policies.
Metrics include AP, AUC, BCE, Brier, 10-bin ECE and full confusion/alert counts.

## Frozen development promotion

Requires all fourteen faculties × five development folds × three declared seeds
× A0,A0M,A1,A2,P1; all principal APs defined; complete comparable operating policies.
Mean seed-averaged cell A2−A1 >=.010; A2−A0 >=.010; A2−A0M >0; A2−P1 >=.005.
At least 3/5 positive fold effects, 8/14 positive faculty effects, 2/3 positive
seed effects for A2−A1. Worst faculty mean A2−A1 >=-.030.
For each beta, mean A2−A1 corresponding threshold F score >=-.010 and absolute
mean alert-fraction change <=.050 across F2–F5. Undefined comparisons fail.

Persist the immutable decision and development-prediction digest BEFORE any
retrospective fitting/evaluation. Integrity preflight may deserialize labels;
retrospective outcome summaries remain quarantined until freeze.
Pass means development research promotion, never deployment certification.
Partial/synthetic/smoke runs return INCONCLUSIVE. A3/P1 cannot replace failed A2.

## Diagnostics and uncertainty

Report all faculty/fold/seed effects, worst harms, training/generalization gaps,
SC/HC/both tensor zeroing and specialized-branch zeroing, latent norms/variance,
positive-rank shifts, threshold/budget rescues and harms, and support of observed
history flags. Joint LayerNorm and the retained full path limit attribution:
interventions are sensitivities, not causal contributions or information removal.

Bootstrap defaults to 500 draws of paired faculty clusters at seed-averaged
cell level, fixed RNG 34029, percentile 95% interval. Retain every observed fold
inside each drawn faculty and preserve defined-cell weighting. Intervals
condition on models/observed periods; they do not characterize future semesters.
Report seed variation separately, without averaging scores into a new ensemble.

Matched HGB is a new reference with source-derived fixed hyperparameters, native
NaN handling and no early stopping. It is not an exact Suite33 replay. Canonical
qte/pte predictions are CacheReference; do not certify Committee identity from
unavailable builder code. Historical fourteen-/thirteen-faculty and IT-only
macro references are metrics-only and carry their population labels.

## Artifacts, resume and software validation

Single log must contain complete tables and automatic restrained interpretations.
Keep identifiers only in local prediction artifacts, not aggregate log tables.
Commit cell fragments atomically after full CPU checkpoint replay (atol/rtol 2e-5).
CPU same-device unit replay uses atol 1e-7. Store all source/configuration/version
hashes; reject changed resumes and regenerate complete final tables/epoch traces.
Resume boundary is a completed cell, not a partially trained epoch.

Validate source contracts, gradients, temporal/target-label isolation, threshold
freeze, macro averaging, corruption rejection, all seven synthetic folds, all
controls, resume, replay, launcher failures, and ZIP integrity. Synthetic scores
are software validation only. Real joblibs and CUDA training need server execution.
