# Suite34 IP-MVCL — finalized design

21 October 2026. The supplied complete design is retained in DESIGN_ORIGINAL.md.
This document resolves its open contracts; it takes precedence where they differ.

The hypothesis, compact residual MLP, full-input path, independent S/C/H encoders,
additive A1, multiplicative A2, capacity control A0M, random-view P1 and optional A3
remain unchanged. One joint logit is trained end to end. No teacher, ensemble,
router, AP loss or sequence pretraining is added.

The full nonlinear MLP and A1 can already learn interactions implicitly.
A2 tests a specific explicit multiplicative inductive bias; A1 is not an
interaction-free function class.

## Binding implementation decisions

1. **Source contract.** Reuse cross_faculty_residual_moe/cache and Suite12
   sidecars. The inspected Suite32 adapter is authoritative: historical rows are
   cf_covered_mask AND finite across every cf_probabilities member. Sidecar
   train_row_index must equal their exact canonical positions, in order.
   Course channels use the inspected twelve static plus five historical fields.
   No newer experiment cache is required. Original joblibs are not present in
   this delivery environment; server preflight remains mandatory.
2. **Schema.** Discover the entire cache population, require seven folds, compute
   one ordered common raw schema and verify per-faculty schema stability across
   folds. A pilot still derives its schema from the entire population.
   The reviewed manifest is an explicit catalog of names observed in original
   source headers and the inspected course sidecar contract. Preflight binds
   that catalog to actual cached names and saves the exact resolved manifest.
   Unknown names remain in the full path. No unverified 136-name list is invented.
3. **Preprocessing.** Training-only median and standard-deviation scaling, fixed
   clipping at 10, one fixed-position cache-nonfinite indicator per original
   feature. All-missing fields use median zero and scale one. Zero-imputed source
   missingness cannot be reconstructed from an already imputed cache. Explicit numeric-coded categorical
   fields retain their raw numeric channel and receive a training-fitted,
   fixed-width vocabulary with an unknown slot. Every arm has identical inputs.
4. **Placebo.** Shuffle original-feature bundles, including all derived channels,
   within equal-width bundle strata. Preserve each view's original-feature and
   transformed-channel counts. Partition is label-free, fixed across all cells
   for a recorded seed. A1/A2 begin from identical weights for a matched seed.
   A0M width is chosen once from the common transformed schema to match A2 within
   5%, then held fixed across faculties/origins.
5. **Chronology.** One canonical history per outer fit. Recompute ordinal as
   year*3+term-1 from target_NHHK; never mix with extracted relative ordinals.
   Start with the latest three global historical semester blocks. If necessary,
   try 2,1,4,5,... blocks, retaining at least one fitting block, until fitting
   contains both classes and validation contains positives and negatives.
   Record every fallback; otherwise fail. Epoch selection averages defined
   validation-semester APs equally. Ties within 1e-6 favor smaller validation BCE;
   exact ties favor the earlier epoch. No random split.
6. **Refit.** Reinitialize the same architecture and seed. Fit new preprocessing
   on all permissible history BEFORE refit. Train exactly the selected epochs.
   Target labels never enter preprocessing, loss, epoch selection or inference.
   F7 may use earlier F6 labels in its fixed historical cache; global protocol
   selection and operating thresholds remain frozen from development.
7. **Thresholds.** Separate faculty×arm×seed policies for beta=1 and beta=2.
   Use only earlier F1–F5 predictions. F1 unavailable. F6 and F7 use the identical
   policy fitted on F1–F5. Maximize pooled earlier-development F-beta; ties choose
   the largest threshold. Unavailable support yields unavailable metrics.
   Fixed budgets select ceil(fraction*n) rows, with row_index breaking score ties.
8. **Retrospective quarantine.** Persist development decision and its prediction
   digest before retrospective fitting/evaluation. Preflight can deserialize
   labels for integrity, but withholds retrospective prevalence/metric summaries
   until that freeze. The decision function accepts development records only.
9. **Promotion.** Full 14×5 development cells, seeds 29/129/229 and all principal
   arms are required. A2−A1 and A2−A0 >= .010; A2−A0M >0; A2−P1 >= .005;
   >=3 positive fold effects, >=8 positive faculty effects, >=2 positive seed
   effects; worst faculty mean A2−A1 >=-.030; mean F1/F2 changes >=-.010 under
   their separate policies; mean alert-fraction changes <=.050 in magnitude.
   Undefined comparisons fail, never silently pass. Incomplete/pilot studies
   are INCONCLUSIVE and cannot promote. A3 is a prespecified secondary extension
   when requested, never an automatic substitute for A2.
10. **Comparators.** Train a new matched HGB with the inspected Suite32 learner
    settings (.06 learning rate, 100 iterations, 15 leaves, 20 minimum leaf rows,
    L2=1, no early stopping). Use identical retained history and raw information;
    retain native NaN handling. Label as a new control, not exact Suite33 replay.
    Cache qte/pte scores are row-aligned within their canonical source container;
    call them CacheReference unless Committee identity is independently certified.
    Published archived macro scores are metrics-only references, never paired.
11. **Diagnostics.** Tensor-zeroing and latent-view-zeroing are sensitivity
    interventions, not causal effects; joint LayerNorm also changes normalized
    contributions. Keep trained A1 as the primary mechanism control. Record
    gradients, activations, positive rank shifts, budget/threshold rescues and
    harms, and support of verified history flags. Unknown history remains unknown.
12. **Uncertainty.** Average matched-seed deltas within each faculty–fold, retain
    all fixed folds, then bootstrap faculty clusters. Report descriptive
    fixed-period intervals and seed SD separately. No row/seed bootstrap may be
    presented as uncertainty over new temporal environments.
13. **Resume.** Source SHA256, package-source hash, resolved manifest and effective
    configuration enter a run signature. Refuse changed runs. Commit each cell
    atomically only after predictions/checkpoint/report fragments are complete.
    Verify all artifact hashes on resume and regenerate the entire final log
    tables from completed fragments. No silent skip by filename alone.
14. **Reporting.** Every cell, arm, seed, diagnostic and comparison is printed
    without truncation; identifiers remain in local prediction files. The
    launcher captures preflight/tests/training/reporting failures in one log and
    propagates exit codes. Smoke/synthetic runs use separate paths and cannot
    produce primary research promotion.

## Validation and practical scope

Deliver source, tests, synthetic generator, checkpoint replay, environment check,
launcher, configuration, protocol and documentation in one top-level directory.
CPU validation is separate from CUDA execution. No real AQ performance is claimed.
The full run is 1,470 neural cell fits, each with selection and refit phases
(2,940 training phases), plus HGB controls and diagnostics. Measure actual runtime
with the real-cache smoke command; synthetic timing is not an estimate.

The user explicitly authorized autonomous design finalization and implementation.
The workspace is a standalone artifact directory, not an external Git checkout.
