**ARCHIVED v1 PROTOCOL.** The stronger origin-verified assumptions below are NOT supported by current AQ exports. For the runnable Suite33 v2 reconstructed experiment and its revised scientific contract, read `REVISED_PROTOCOL.md` first. Certified-mode source gates remain separately available.

---

# AQ research after Suite32: an origin-verified offering-context hypothesis

**Research and design freeze:** 17 October 2026  
**Status:** Research protocol; no Suite33 training or outcome generated.  
**Proposed suite:** Suite33 — Origin-Verified Offering-Context Learning (OVOCL).  
**Primary question:** Does the historical preparation of the other students actually enrolled in the same course offering, as known at semester origin, improve *this student's* course-failure ranking over the strongest own-history/course representations?

## 1. Decision

Prioritize one new information source—not another complex model, faculty clustering scheme, or ensemble: **the origin-time learning environment of an individual course offering**. Construct observable, leave-one-student-out summaries of peers' *past* academic histories using an independently verified enrollment roster. Feed these along with the target student's already established own-history and target-course predictors to **one jointly fitted classifier**. Study the added information through carefully matched course-wide, section-specific, and section-placebo contrasts.

This is not claimed as a previously undiscovered scientific method. The distinctive hypothesis is a new, falsifiable *AQ information contract*, whereas related grade-prediction and randomized peer-effect work predates this study. It is a predictive context hypothesis, **not** proof of causal peer influence.

**Non-negotiable gate:** The extracted outcome-filtered test CSVs are insufficient to certify enrollment membership at semester origin. We will not call a result “semester-origin leakage-safe” without enrollment snapshots or event reconstruction. If independent rosters cannot be verified, the candidate remains an explicitly exploratory post-extraction analysis, not a promoted framework; the prerequisite data-contract audit becomes the only valid next step.

## 2. Evidence and why this is a change of direction

### 2.1 Historical results that constrain the project

- Suite25 Committee blend: F6/F7 14-faculty macro AP **0.369526**; Suite22 safe blend **0.369159**. These are composites, not standalone models.
- Committee reference: **0.350152**; original Council's archived 13-faculty macro **0.364204** (English absent); Council's **0.444927** is an IT-only result and must never be compared with 14-faculty macro.
- Suite31 full-history HGB **0.353807** (different training contract); Suite32 local common HGB **0.346168**; Suite32 selective common **0.345592** and selective native **0.339950**. Suite32's policy admitted borrowing in only 24/98 cells and its F6/F7 admitted cases showed weak and inconsistent transport.
- Suites29–30 model *different courses taken by one student concurrently*. Suite25 models *previously graded registrations*. Suite32 changes *labeled historical training donors across faculties*. None of these is the proposed **other currently enrolled students' earlier histories**. Source-code inspection of Suite30 confirms groups are keyed by `(faculty, student_id, target_semester)`; its “peers” are other course tokens belonging to that student, not classmates in the same section.

### 2.2 Direct data facts in the supplied 17 October audit

- The seven test folds contain **324,420** canonical numerically graded registrations and **12,590** failures. The evaluation target is grade `<5`, at semester origin.
- F7 extracted tables contain **53,648** total rows organized into **1,639** `(semester, course_id, section_id)` groups, median size **32**. **52,654** rows appear in groups of at least 10. **444** sections cross faculty boundaries and **22,659** of the extracted F7 rows belong to multi-faculty offerings. These are *final extracted outcome tables*, not certified as-of rosters.
- A leave-one-student-out prior-GPA peer mean has approximately **0.526 SD** and **0.373 correlation** with one's own previous GPA in the reported F7 audit. This is evidence of variation and possible redundancy, not incremental AP.
- F6 failures account for **5.5237%** of canonical target registrations; F7 failures **7.7901%**. Among failures, grades in `[4,5)` grow from **48.64%** at F6 to **53.12%** at F7; zeros shrink from **16.66%** to **9.69%**. Earlier F1 grades `[4,5)` account for **24.38%** of failures, zeros **34.63%**. The failure phenotype and prevalence shift, but causes are unidentified.
- All faculty-specific raw columns beyond the 136 common raw columns have a `course_status` prefix in inspected F7 source schemas. The 153-feature baseline consists of **136 shared raw + 17 structured course features**. A universal “return to 275” is therefore not justified; 275 is IT-specific and these private columns are primarily transcript-status vocabulary.
- F6 original graded training tables contain **480,967** rows / **10,897** failures, versus Suite32 retained **470,165** / **10,533**; the F7 discrepancy is the same **10,802** rows / **364** failures. Exact identity, eligibility, availability and alignment must be audited. Stable aggregate differences do not prove valid extra examples.
- Original split values contain blanks as well as genuine zero values; completed caches have masked this distinction. A naive zero-as-missing indicator did not help in Suite32, but did not test pre-imputation missingness. Historical numeric preprocessing may use train-column means and zero fallback; do not assume blanket zero imputation.

### 2.3 Why another model architecture is lower priority

V2/Council, RF/ET/HGB, MLP, tree–memory combinations, TabM/AGOP/ModernNCA, relational feature graphs, structured grades/readiness, SACR historical retrieval, within-student set transformers, TabICL/TabPFN and donor gating have already been investigated. The repeated problem is weak out-of-time realization of optimistic historical relationships. A new complex learner with the same input cannot restore distinctions never supplied at training and deployment.

## 3. Ranked hypotheses and their discriminating questions

| Priority | Hypothesis | Distinct new observation? | Main uncertainty | Decision |
|---|---|---|---|---|
| Gate 0 | Recover legitimate training coverage and original missingness states | Yes, at data-pipeline level | Whether excluded rows are truly eligible/available; whether missingness adds value | Mandatory bounded audit, not a thesis-level new architecture |
| 1 | **Current course-offering incoming composition** | **Yes:** other current classmates' *prior* histories | Were their identities/measurements known at origin? Redundancy with own history and course/course-cohort sorting | **Selected primary direction** |
| 2 | Current instructor/assessment-rule version | Yes if verified source versions exist | No confirmed valid as-of metadata in split archive | Reserved conditional extension, not inferred from outcomes |
| 3 | Verified joint transcript relationships lost by aggregation | Possibly | Strong historical relational models may already encode them; meanings of status codes not established | Only a source-level counterexample warrants an arm |
| Diagnostic | Temporal change in failure grade distribution and outcome selection | Observed, but not automatically a new origin predictor | Unknown grading/selection mechanisms | Diagnose shift; preserve fixed `<5` target |
| Defer | Faculty clusters, external FM, generic class rebalance, another MoE, larger same-student portfolio | Mainly reused existing evidence | Negative historical transport and overlap with exhausted experiments | Not Suite33 |

## 4. Formal problem and distinction from past suites

For each eligible target registration `i = (student s, course c, section k, semester t)`, set:

`y_i = 1[grade10_i < 5]`, where `grade10_i` is the eventual eligible numeric outcome.  
`X_i(t-)`: student and course features that truly existed and were released before origin `t0`.  
`R(c,k,t;t0)`: the roster of unique students enrolled in this offering **as actually known at t0**, including students whose eventual graded eligibility is absent.  
`H_j(t-)`: an admissible summary of classmate `j`'s released grades and prior academic history before `t0`.

For each target student, construct a **leave-one-student-out** context:

`P_i = aggregate({H_j(t-): j in R(c,k,t;t0), j != s})`.

Also form `G_i`, the corresponding *course-by-semester* context pooled over all sections, excluding the target student. Compare the section profile to the course-wide profile, rather than treating all section composition as evidence of a separate peer effect.

**Primary unified function:**

`p_i = sigmoid(F_theta(X_i(t-), target_course_i(t-), G_i, P_i, own_relative_i))`.

`F_theta` is **one jointly trained prediction function per target faculty**, using one compatible learner recipe across faculties, one loss and no Committee/SACR probability blend. Separate per-faculty fits respect heterogeneous faculties; other faculties can contribute *unlabeled origin-time peer attributes* when they share a section, not donor target labels. A cross-faculty global learner is **not** the hidden primary variant.

### Novel distinction

- Suite25: retrieve *past graded examples* (target-relative historical memory).
- Suite29/30: model *one student's concurrent course set* and joint failure events.
- Suite32: *change the supervised historical source population* used to train a faculty learner.
- Suite33: describe *the students sharing this current offering* using *their earlier histories*, without their current outcomes.

A negative donor label-transfer result from Suite32 does not rule out this form of contextual information.

## 5. Gate 0: canonical row/data audit (mandatory, bounded)

This must be completed before attributing gains to offering context; it is not a new model-search campaign.

1. Reconcile stable registration keys (student, course, semester, registration ID when available), labels, and metadata between original split CSVs, canonical train/test caches, the cross-fitting coverage mask, and the structured 17-feature sidecars. Explain all **10,802 / 364** historical differences by identity and exclusion reason, not aggregate counts alone.
2. Prove historical label *release time* (not just course semester) is earlier than target prediction time for every training label/feature. If publication time is missing, explicitly state where the existing semester proxy is relied on; do not certify strict availability.
3. Inventory field types: counts, genuine continuous quantities, ordinal encodings, nominal IDs, source missing values, not-applicable fields, and imputed fields. Confirm definitions of `course_status` 0/1/2 before using them as semantic states.
4. Test coverage and missingness in a small 2x2 **raw-feature-only** control: covered vs additional *verified eligible* training rows, and old numeric preprocessing vs a learner that handles original blanks. Same target rows, learner capacity, label chronology and raw feature subset.
5. Rebuild the 17 structured features only if extending the historical training universe; never index the existing sidecars by unverified extra rows. The principal offering-context experiment may instead keep the **original canonical covered training universe** to preserve a faithful S0 control.
6. Freeze the row identities and imputation contract before any context experiment. Stop Gate 0 after the planned controls; do not launch arbitrary imputation architectures following a null result.

Outputs: row reconciliation table, leakage/availability audit, missingness inventory and four-cell metric table. If this audit changes the valid training contract, rerun S0 on exactly that contract and compare context arms *within it*.

## 6. Gate 1: origin-valid offering roster (critical)

### Required input

Historical enrollment events or dated snapshots for **every semester used as a training, development or retrospective target**. Minimum fields: globally stable student key, semester, course ID, section/study-group ID, enrollment status, observed/known-from timestamp, effective/event timestamp and snapshot cutoff. A latest-export registry without version history may not suffice. Confirm whether section IDs can be reused and whether `IDToHoc` indeed identifies the operative teaching group; map cross-listed course identities carefully.

Construct `R(c,k,t;t0)` from event/snapshot states **known at the exact same institutional origin** as the target `X_i`. Use `(semester, course, section)` as the offering key; never section ID alone. Keep all active/registered classmates at the origin, whether they later withdraw, receive no grade, or fall outside benchmark eligibility. Exclude explicitly not-yet-registered, canceled or later-entered enrollments according to the documented snapshot policy.

### Reject these shortcuts

- Computing peers solely from the eventual numeric-grade test rows.
- Using a final roster with late adds or post-origin deletions, even if its numeric attributes are previous GPAs.
- Using target or peer *current semester* grades, eventual pass rates, attendance, withdrawals or eligibility to decide inclusion.
- Joining students across faculties with accidentally faculty-local identity keys.
- Treating a section identifier as a continuous numerical score or historical difficulty estimator.

### Proof-of-availability audit

For every origin: report number of target rows, targets matched to an origin snapshot, distinct offerings, roster members, cross-faculty offerings, historical-data coverage, event time violations, duplicate memberships and deviations from final extracted rosters. Preserve unresolved cases as missing context with explicit flags, never silently replace with final-roster data. Student-level previous-GPA measurements must carry a release cutoff where verifiable. Training origins must be reconstructed independently of validation outcomes.

**Stop rule:** If genuine as-of rosters cannot be reconstructed for an analytically meaningful part of the population, the peer-context result is only exploratory; do not run it as a certified leakage-safe primary Suite33. Document the missing source-data contract rather than weaken it to obtain a score.

## 7. Origin-context feature specification — fixed before outcomes

Use the existing **153-feature S0 input** as the main substrate. Do not universally append private `course_status` columns. The group information must be computed from raw, origin-valid histories and roster identities, not from fitted risk scores or current labels.

### Feature family A: size, observation and support (S1)

For target `(s,c,k,t)`, where `R` contains unique students at origin, calculate:
- `peer_count = |R\{s}|`, `observed_peer_count` for each historical measurement and separate fractions missing/released;
- section total size, distinct faculty count, origin roster completeness, previous GPA availability fraction;
- optional number with prior semesters and course-history support, if source semantics are confirmed.

Counts and missing support remain explicit. A peer mean based on 3 students is not equivalent to one based on 40 students. Do not interpret source blanks as zero GPA.

### Feature family B: course-wide incoming history (S2)

Within `(semester, course)` across **all sections**, aggregate peers' origin-known earlier GPA and verified history variables, excluding the target student. Prespecified small block:
- count with historical GPA; mean, SD, and a modest range/quantile summary of prior GPA;
- fraction below a previously fixed GPA threshold and fraction with verified warnings, if meaningful;
- source observation fractions.

This is a control for course-by-cohort intake composition. It can carry signal without any section-specific composition effect and is distinct from historical *outcome* difficulty.

### Feature family C: actual section incoming history (S3)

Apply identical summaries to `(semester, course, section)` peers excluding target `s`, and add section-minus-course mean and dispersion contrasts. Retain the component support counts and section-count controls. Small sections with few observed peers receive missing estimates plus observed-support flags (or a **prespecified**, label-free shrinkage-to-course-mean rule); do not pretend an imputed/shrunken mean was observed.

### Feature family D: target-relative location (S4, selected primary representation)

- Own previous GPA minus peer mean, with coverage and source-quality flags.
- Optional own GPA percentile among observed historical classmate GPAs with a missing indicator; compute from the *same origin time* and ensure the student's own GPA does not leak into the peer aggregate.
- Small interaction-ready contrasts between target preparation and group variability; the learner is responsible for estimating interactions from development data.

Own-relative features are transformations of existing own/G/P information; they are not counted as an independent source of information. No protected-category group features are introduced. Suppress/smooth very small group summaries as appropriate for sensitive records.

### Computation

Build group-level sums, counts, squared sums and sorted value buffers once per `(semester, course, section)` and `(semester, course)` to compute leave-one-out values without O(group_size^2) loops. Exclude the student at both group levels. Compute quantile/percentile algorithms deterministically. Persist origin, roster snapshot version, eligible input keys and a provenance hash. The primary output is one feature row per original benchmark target, never a changed target population.

## 8. Predeclared experiment matrix

All arms use exactly the same target labels, target rows, origin timetable, own-history baseline, learner family, training budget and preprocessing, with only the declared feature block changing.

| Arm | Inputs | Question |
|---|---|---|
| **S0** | Strong 153-feature own-history/target-course HGB | Reference, reproduce close to Suite32 local-common under a frozen contract |
| **S1** | S0 + section size, observation/coverage only | Is the signal just section size or whether histories exist? |
| **S2** | S1 + **course-wide** incoming preparation | Is course-cohort composition useful without section-specific detail? |
| **S3** | S2 + actual section leave-one-out preparation | Main genuinely section-specific information contrast |
| **S4 (primary)** | S3 + own-relative preparation contrasts | Can one jointly fitted model exploit context conditioned on individual preparation? |
| **P1 placebo** | S4, but section composition replaced by assignment matched within `(course, semester)` while retaining counts, own predictors and course-wide values | Is actual section assignment more informative than generic course intake? |

Only actual metadata and historical preparation are used; no peer current outcomes. Placebo swaps maintain section-size and source-availability strata and, where possible, faculty/cohort composition. Only courses with at least two interchangeable sections contribute to a clean section-assignment placebo. Record matched coverage; do not extrapolate conclusions to unmatched courses. Include a sensitivity against random section membership that is independent of target outcomes. Do not use the placebo to search for the best of many generated shuffles.

**Primary information test:** `(S3-S2)` for section specificity and `(S4-S1)` for the total preparation-plus-relative increment. `S4-S0` is the broad system uplift but not sufficient for mechanistic attribution. P1 distinguishes actual section associations from a generic course/cohort effect, not a formal causal estimate.

## 9. Predictor, training and selection

### Default reference learner

Begin with the verified strongest compatible HGB baseline under the same source-feature and training-row contract, including a reproduction of Suite32's local-common recipe. The **final S4 is a single HGB model**, not late blending, and does not consume Committee or Council probabilities. Fit per faculty using an identical protocol across fourteen targets. Train on target-faculty labeled registrations, while computing peer *attributes* globally from all known origin classmates. No source-faculty training labels are borrowed by default.

Do not treat HGB=the universal best model: Suite31's full-history HGB reaches 0.353807 under a different recipe. Benchmark the richer historical HGB recipe **separately with row and feature alignment**, and include Committee/Council and Suite25 blend as strong performance references.

### Predeclared interaction-capacity check (not an architecture fishing expedition)

If development screening finds repeatable `(S3-S2)` or `(S4-S1)` information but the simple HGB interaction is weak, allow **one** GPU-enabled CatBoost contextual model as a confirmatory matched-input test. Evaluate CatBoost-S0 versus CatBoost-S4 with the same numerical/categorical schema. This comparison tests learner integration, not merely whether CatBoost alone is stronger. CatBoost supports `task_type='GPU'` on CUDA hardware but its GPU implementation can be non-bitwise-deterministic; pin version, devices and seeds, and test seed stability. CatBoost is *not* part of the primary Suite33 hypothesis by default; do not mix its predictions with HGB.

### Selection schedule

1. Recover and freeze the origin-aware row/feature contract; do not condition inclusion on eventual target eligibility for *peers*.
2. Construct every historical training origin's peer features as-of its own cutoff, not by looking at a later fold's roster/history.
3. Reproduce S0; evaluate S1, S2, S3, S4 and prespecified P1 on F1–F5 using expanding chronological outer folds.
4. Within each development origin, only use earlier origins to select learner iterations/regularization and **F1/F2 thresholds**, with explicit abstention/fallback for earliest origins without threshold-fitting support. Do not tune thresholds on the same period whose F1/F2 is being evaluated.
5. Freeze one selected arm/learner and thresholds using F1–F5 only; accept or reject peer information on the preregistered criteria below.
6. Run a **single F6/F7 retrospective comparison** with frozen decisions. F6/F7 are not pristine; no iterative design changes should be claimed as prospectively confirmed on them.
7. Request a later held-out semester if available for genuinely new evaluation.

### Fallback rule

If S4 fails the fixed development promotion gates, **do not train an invented mixture** or pick the best F6/F7 arm. The selection result is `S0` or the strongest prespecified supported simpler arm under the agreed rules. Report a negative Suite33 result without overstating a marginal gain.

## 10. Why we might observe improvement — and why we might not

### Hypothesis H1: course-cohort selection

Even before section assignment, course-specific contemporaneous intake ability distribution may help distinguish registrations with identical own historical summaries. Expected pattern: S2>S1; S3≈S2; matched section placebo close to true context. This would be a useful *course-environment* predictive feature, not evidence of section-specific peer effect.

### Hypothesis H2: section assignment and relative preparation

Within the same course and semester, peers' historical preparation and dispersion may provide context not contained in course-wide means. Expected: S3>S2 especially for multi-section courses, and actual assignment outperforms section-membership placebo. Additional S4 gain may be concentrated in students far above/below their section's incoming GPA profile. Selection into classes remains a possible explanation.

### Hypothesis H3: availability/size artifact

Section size and whether peers have historical GPA could be the only useful component. Expected: S1 gains, S2/S3 little extra; or placebo performs equally well. If so, the mechanism is not demonstrated and should not be advertised as classmate compatibility.

### Hypothesis H4: confounding and redundancy

The rich 153 baseline already contains enough course, student, difficulty and cohort predictors; peer GPA mostly follows own GPA (observed F7 correlation ~0.373) and faculty/curriculum assignment. Expected: S2/S3/S4 gains vanish on strong own-history controls or when restricting to multiple-section courses.

### Hypothesis H5: true origin roster unavailable

Final roster membership reflects later enrollment changes/outcome eligibility. Apparent improvement may be inflated by time-inappropriate selection. If provenance fails, these models can still be run only under a clearly labeled **exploratory final-roster upper bound**, never as evidence of deployable semester-origin ranking.

### Hypothesis H6: assessment regime

Later grade distribution and positive-case mix suggest a changed outcome-generating environment. Verified historical instructor or assessment-rule versions might add independent conditional risk, but the present split data do not contain a certified as-of versioned source. Keep this as a separate future data acquisition hypothesis, not a new decoder or a cause inferred from grade distributions.

## 11. Evaluation and falsification

### Primary target and macro score

Use the canonical registered-course outcome `grade10<5` on the existing eligible graded target rows. Equal-weight macro AP across **70 F1–F5 faculty-fold cells** and **28 F6–F7 cells**, with exact target IDs. Do not mix Council's 13-faculty archive with 14-faculty numbers; show a separate support-matched Council comparison. Show fold and per-faculty AP, AUC, Brier, ECE, F2@20 and all primary arms.

### Operational thresholds

Select F1-optimal and F2-optimal thresholds on earlier development evidence only; freeze a selected F1–F5 policy for F6/F7. Report P/R/F1/F2 and confusion matrices at both thresholds, flagged unique students, flagged course registrations and alerts per 1,000 rows. Fixed 20% budget remains an additional, not replacement, metric.

### Uncertainty

Save predictions with stable global student and offering keys. Resample students (including all their registrations and faculty memberships) to construct descriptive paired faculty-fold macro-AP intervals; additionally cluster by offering for a sensitivity analysis, documenting that separate clusters are not a formal crossed-cluster interval. No claims of guaranteed next-semester improvement follow from historical bootstrap intervals. If a bootstrap cell lacks one class, predeclare a consistent method and report invalid resamples instead of changing cell weights without notice.

### Heterogeneity and failure phenotype

Analyze gains by section size, valid peer support, number of faculties in offering, seen/unseen course, one versus multiple sections, target own-GPA bands, relative-gap bands, new versus returning students, course-history flag, grades `0` versus `[4,5)` **for post-hoc diagnostics only**, and faculty/semester. Never change the target or train a grade-zero administrative subtype without verified labels.

### Practical promotion gates (proposed, not previously demonstrated facts)

- **Gate 1**: independently verified as-of rosters and release cutoffs; no target-outcome-based roster selection.
- **Evidence gate**: positive S3-S2 or S4-S1 in several development origins; no gain only from roster-size/coverage.
- **Placebo gate**: S4 beats P1 on matched multi-section support, otherwise label gains as course/cohort association instead of section assignment signal.
- **Breadth gate**: gains not dominated by one faculty, one origin or a handful of failure events; report wins/losses/worst loss and negative subgroups.
- **Promotion-size signal**: illustrative `>= +0.010` absolute F1–F5 macro AP over *strongest matched information baseline*, accompanied by positive temporal stability and acceptable operational metrics; below that, keep result as a minor local finding, not substantial advancement.
- **Retrospective benchmark**: F6/F7 competitive with Committee **0.350152**, Suite25 blend **0.369526**, Council 13-faculty **0.364204**, and IT-specific Council **0.444927**. These comparisons require aligned support caveats. No guarantee is asserted.

Stopping is an acceptable scientific result. Do not revise the protocol because F6/F7 happens to favor one arm.

## 12. Training/runtime, logging and saved artifacts contract

### One command, one log

The actual later implementation shall expose **only one normal launcher** for the full run:

`./run_suite33.sh`

It performs environment and data gates, feature construction, S0–S4/P1 development, development freeze, F6/F7 retrospective evaluation, reporting and exit-status verification in order. Optional internal resume flags are developer conveniences, not multiple required user invocations.

**One consolidated log:**

`/root/disk_3tb/earlywarning/30072026/log/suite33_offering_context.log`

All stdout/stderr, fold/faculty results, gate decisions, feature/roster provenance, training curves/costs, alerts, insights, warnings and tracebacks are collected there with timestamps and a run ID. Do not generate separate stage-specific logs by default. A failed gate yields a nonzero exit code and an explicit explanation in the same log.

**All results:**

`/root/disk_3tb/earlywarning/30072026/results/suite33_offering_context/`

Suggested outputs:
- `manifest.json`, dependency versions, code hash, data-roster snapshot hashes, target and row counts;
- `audits/` source column semantics, coverage reconciliation, imputation provenance, as-of roster proof, cross-faculty section identity, source release cutoffs;
- `context/` content-addressed origin peer feature caches, section/course aggregate metadata and missing-state ledgers;
- `development/` full 70-cell model/AP comparison and prequential threshold selection, locked decision manifest;
- `retrospective/` full 28-cell model/AP comparison and original Council support alignment;
- `predictions/` row-keyed, fold/faculty-identified scores for every arm and every original target;
- `insights/` S4-S0/S4-S1/S3-S2 paired deltas, placebo-matched scope, section-size strata, own-relative strata, features' conditional ablations, sensitivity intervals, risk-profile drift, computational cost;
- `reports/` detailed score tables and an interpretable hypothesis checklist with reasons for success or failure.

### Computation

HGB runs on CPU; do not suggest GPU failure because CUDA utilization is zero. A separately permitted CatBoost integration test uses NVIDIA CUDA (available V100 previously) after a verified `task_type=GPU` environment check, while documenting known GPU nondeterminism. No architecture should be changed silently to maximize GPU utilization.

### Required print records

At minimum `[START]`, `[SCIENTIFIC_CONTRACT]`, `[DATA_GATE]`, `[ROW_COVERAGE]`, `[ASOF_ROSTER_AUDIT]`, `[OBSERVATION_PROVENANCE]`, `[SECTION_COHORT_AUDIT]`, `[PEER_FEATURE_BLOCK]`, `[FOLD_START]`, `[TRAINING_DIAGNOSTICS]`, `[FACULTY_RESULTS]`, `[CONTEXT_INCREMENT]`, `[COURSE_VS_SECTION]`, `[PLACEBO_MATCH]`, `[RELATIVE_GAP]`, `[FAILURE_MIX]`, `[THRESHOLD_FREEZE]`, `[RETROSPECTIVE_HISTORICAL_COMPARISON]`, `[DECISION]`, `[INSIGHT]` and `[RUN_STATUS]`. These are logging requirements, not claims that code already exists.

## 13. Explicit boundaries and future extensions

1. **Instructor / rules:** a second-stage verified offering metadata block only if official as-of assignments, assessment rule versions and earlier released outcomes exist. Compare metadata support versus actual metadata, seen versus unseen instructors, and rule-change strata. A numeric section ID is not a teaching-effect surrogate.
2. **Transcript joint records:** only after establishing a pair of real admissible histories collapsed by the strongest current representation, with verified codes and target-relevant course mappings. Do not retest generic readiness, more GNN/Transformer layers or student-course retrieval under a new name.
3. **Outcome mechanism:** use the shift toward near-pass failures for shift diagnostics; do not call grades 0 “withdrawal” or relabel graded cases by post-hoc guesses. Any two-stage outcome decomposition would need independently documented administrative labels and a predeclared origin task.
4. **Institutional source data:** if the origin roster/rule history cannot be retrieved, improving source collection may have a higher expected research value than trying increasingly complex models on an unchanged cache.
5. **Novelty statement:** to claim novel published research, the contribution would need to be the AQ-specific validated information construction, its origin-time leakage defense and controlled evidence—not the invention of aggregation or the peer-effect hypothesis.

## 14. Final decision statement

**Freeze OVOCL-S4 as the primary scientifically interesting candidate, contingent on source-valid enrollment rosters; run Gate 0 before any new large model.** The experiment is designed to be rejected if course-wide composition, coverage, or label-ineligible final-roster effects fully explain the apparent gain. S3-S2, S4-S1 and the course-matched section placebo are compulsory mechanism tests. Successful ranking must be evidenced on matched folds/faculties over a strong single-learner control, then compared transparently with Council/Committee/Suite25. If source validity cannot be established, do not silently substitute an easy-to-run but invalid peer-graph model.

## Reference bibliography (primary and source record)

- User-supplied `AQ_after_Suite32_data_nature_and_new_horizons(1).md`, 17 October 2026. It summarizes its own audits of `training_data.zip`, AQ code, Suite32 logs, and historical suite reviews. Counts repeated here are source-reported and not independently rescanned in this design session.
- Gardner, Brooks & Li (2018), *Learn From Your (Markov) Neighbor*, *Journal of Learning Analytics*, 5(3), 42–59. https://learning-analytics.info/index.php/JLA/article/view/6146
- Booij, Leuven & Oosterbeek (2017), *Ability Peer Effects in University: Evidence from a Randomized Experiment*, *Review of Economic Studies*, 84(2), 547–578. https://academic.oup.com/restud/article-abstract/84/2/547/2417357
- Ong, Zhu & Mossé (2022), *Towards Including Instructor Features in Student Grade Prediction*, EDM 2022. https://educationaldatamining.org/edm2022/proceedings/2022.EDM-long-papers.21/
- Carrell & West (2010), *Does Professor Quality Matter?*, *Journal of Political Economy*, 118(3), 409–432. https://www.nber.org/papers/w14081
- Fauszt (2026), *Time-Consistent Prediction in Higher Education*, *Information*, 17(6), 581. https://www.mdpi.com/2078-2489/17/6/581
- Forbat & Gemulla (2026, preprint), *RelICL: Training-free Relational Learning with Tabular Foundation Models*. https://arxiv.org/abs/2610.01725
- Huang et al. (2026, preprint), *RelAgent: LLM Agents as Data Scientists for Relational Learning*. https://arxiv.org/abs/2605.07840
- Official CatBoost documentation: GPU processing units and nondeterminism. https://catboost.ai/docs/en/references/training-parameters/processing-unit and https://catboost.ai/docs/en/features/training-on-gpu
