**Suite28: ability–requirement research and design decision**

Research date: 7 October 2026. This is a research proposal and architecture specification. It contains no measured Suite28 results and no implemented Suite28 training package.

**Recommendation**

Retire direct tree–memory co-adaptation as the main research direction. Investigate a **Dynamic Readiness–Requirement Model (DRRM)**: infer a compact, uncertain student state from chronological course outcomes, interpreting each outcome relative to the course's estimated requirements, and use that state to predict the next semester's courses. Rich current features inform the state forecast. Historical outcomes update it through the same outcome likelihood used for prediction.

Use a lean, contextual dynamic multidimensional IRT model as a compulsory comparator. Keep a flexible transcript-plus-tabular model as a separate alternative if the structured predictor discards useful information. These are three different modeling commitments, not three modules to combine immediately.

The research question is whether **difficulty-aware inference and a course-specific multidimensional state improve future ranking**, especially when histories are sparse or uneven. There is no evidence yet that this will beat Suite25 blend, and no theoretical argument guarantees a large AP gain. The proposal is worthwhile because it tests a different representation and learning mechanism with a much smaller computational commitment than Suite27.

**1. What our experiments actually establish**

The completed Suite27 review remains the empirical starting point. The following numbers are extracted from its verified final-log tables. The historical holdout consists of F6 and F7: 97,004 course rows, 6,574 failures, 28 faculty–fold cells, and 12,287 distinct students. Primary AP is the unweighted mean of eligible faculty–fold AP values, not AP on concatenated rows.

| Model | Development macro AP | Historical holdout macro AP | Holdout coverage |
| --- | ---: | ---: | ---: |
| Suite25 blend | 0.369071 | 0.369526 | 28/28 cells |
| Suite22 | 0.373747 | 0.369159 | 28/28 cells |
| Suite26 MINT | 0.368789 | 0.338258 | 28/28 cells |
| Suite27 MCGB | 0.426255 | 0.341348 | 28/28 cells |
| Suite27 deterministic memory | 0.425759 | 0.341715 | 28/28 cells |
| Suite27 current-only stock | 0.304560 | 0.322058 | 28/28 cells |

Suite27 MCGB loses the two-fold faculty-mean AP comparison to both Suite22 and Suite25 blend in all fourteen faculties. Its advantage over MINT is only 0.00309052 AP; the reported paired student-cluster interval crosses zero. Its deficit to blend is 0.02817731 AP, with the reported interval entirely below zero. Its small deficit to deterministic memory is also uncertain. These are comparisons of the supplied fitted checkpoints, with one Suite27 seed; they do not cover all training or future-semester variability.

The most relevant mechanism evidence is more specific than “memory does not work.” The deterministic history package gives most of Suite27's gain over its own current-only controls. Learned relevance, path feedback, and co-adaptation add little reliable marginal holdout value. In F7, masking explicit memory in the already fitted MCGB forest raises AP from approximately 0.317875 to 0.350803. That is a postfit diagnostic, not the result of retraining a no-memory model. The reversal also occurs in nonempty-history strata; it cannot be attributed entirely to missing histories.

The fraction of target rows with zero eligible history grows from approximately 0.79% in F6 to 27.70% in F7. Zero available history is not automatically a genuinely new student: missing transcript coverage, ID alignment, and eligibility rules can produce that condition. Approximately 15.58% of F7 query courses are unknown to the Suite27 training course vocabulary. These two forms of incomplete support require different handling.

Suite27 achieved the intended internal tree interaction. Its failure is not evidence that the implementation secretly remained a posthoc calibrator. It is evidence that the tested interaction machinery did not produce a competitive, temporally stable predictor. A redesigned Suite28 should preserve useful chronological information without carrying that computational machinery forward.

The complete prior comparison and source-linked audits are in `suite27_full_review/Suite27_complete_experiment_review.md`. This proposal does not invent results for versions or faculties without aligned measurements.

**2. Distribution shift is a concern; prevalence alone is not the explanation**

The F6/F7 pooled failure prevalence rises from approximately 5.52% to 7.79%. Cohort composition, course mix, history support, score distributions, and feature provenance also deserve attention. The logs do not isolate their causal contributions.

Under pure label shift, the class prior changes while the class-conditional feature distributions remain fixed [R24]. For a fixed scoring rule and threshold, write prevalence as π, true-positive rate as TPR, and false-positive rate as FPR. Then:

$$
\mathrm{Precision}=\frac{\pi\,\mathrm{TPR}}{\pi\,\mathrm{TPR}+(1-\pi)\,\mathrm{FPR}}.
$$

Holding the score distributions within each class fixed, increasing π increases precision at a fixed recall, and therefore increases population AP. This is an analytical consequence of the formula, not an empirical explanation of our folds. Finite samples and changing conditional distributions need not follow that simplified comparison.

A common logit-intercept or prior correction preserves ordering within a faculty–fold cell. It therefore cannot improve that cell's AP. Different course-specific corrections can change within-cell ranking, but must be estimated from legitimately earlier information. Consequently, “handle prevalence shift” is not a sufficient mechanism for beating our primary AP benchmark.

The useful hypothesis for Suite28 is that outcomes from courses with different demands should not be treated as interchangeable evidence about a student. If that representation transfers better across course mixes and student populations, ranking may improve. This is a hypothesis about conditional structure, not a claim of automatic shift invariance.

F6/F7 have already informed several designs. They can still serve as historical comparisons after freezing a recipe, but they cannot become a newly untouched confirmation set. A later unseen semester would be the strongest confirmation if available.

**3. The earlier ability–requirement proposal needs correction, not a new label**

The recovered earlier proposal and the original Suite26 TAIR source already contain a multidimensional ability–requirement head. In the original source, the core score is:

$$
z=\frac{a_c^\top(\theta_s-r_c)}{\sqrt{K}}-d_c.
$$

Three distinctions matter before calling Suite28 a fresh test.

First, that score was trained with the failure label, although its positive discrimination means that larger θ raises the score. As an “ability” interpretation, the sign is backwards: higher readiness should decrease failure probability while the course and context are fixed. A freely learned vector could still behave as a risk representation, so the sign issue does not prove predictive failure by itself. It invalidates the claimed proficiency meaning.

Second, expanding the equation gives a student term and a scalar course term:

$$
z=\frac{a_c^\top\theta_s}{\sqrt{K}}-\left(\frac{a_c^\top r_c}{\sqrt{K}}+d_c\right).
$$

Thus a linear scalar prediction cannot separately identify the requirement vector and free course difficulty. The data only see their combined scalar offset. This is an exact algebraic redundancy, not a training-quality issue.

Third, the earlier ability vector was built with target-dependent attention. The primary predictor remained a bounded modification of the tree score, and unrestricted interaction pathways could bypass the ability head. An auxiliary ability loss therefore did not establish that a coherent ability state drove prediction. The original requirement encoder also omitted a course-ID embedding, relying on course metadata alone.

For Suite28, the student state should be shared across candidate courses within the same prediction episode. Course requirements belong to the decoder. The structured branch should predict outcomes directly, with a correctly signed likelihood. A separate flexible predictor is an explicit alternative/control, rather than a hidden bypass used to claim that the structured theory succeeded.

Source evidence: `suite26_review_input/extracted/suite26_tree_native_student_memory/suite26_model.py` and `train_suite26_tair.py`, alongside `Suite26_design_review.md`.

**4. What the ability–requirement idea can offer**

**Difficulty-adjusted evidence.** Failing an unusually difficult course should not mean the same thing as failing an easy course. Conversely, passing a difficult course can provide stronger evidence than passing an easy one. In a simple correctly signed logistic model, the observed-outcome gradient naturally implements this distinction. A failure with predicted failure probability 0.05 gives a much stronger negative readiness update than an expected failure with probability 0.70, holding discrimination fixed. This toy calculation explains the mechanism; those probabilities are illustrative, not measurements from our dataset.

**Course-specific strengths and weaknesses.** Two students can have similar overall GPA while differing in the dimensions that matter for the next course. A multidimensional state can represent that distinction. Whether our outcomes identify more than a general performance factor is unknown; increasing K does not manufacture information.

**Support-aware confidence.** Ten courses measuring nearly the same direction do not necessarily identify all four readiness dimensions. A covariance matrix can preserve uncertainty in poorly measured directions. Counts and recency alone cannot express this geometry.

**Inductive inference.** A new student does not need an optimized student-ID parameter. The model can initialize from available covariates and update when permitted outcomes arrive. This is established statistical/ML practice, not a novel Suite28 invention [R11, R14, R17, R18].

**A clearer learning objective.** Student inference and course requirements are trained to predict genuinely subsequent outcomes. The framework no longer depends on an auxiliary reader that must also happen to supply useful custom tree splits.

**A compact computational state.** A four-dimensional mean and covariance can be carried through semesters once per student. Multiple target courses reuse it. There is no requirement to run a history reader at each tree node or repeatedly regrow forests.

These strengths are inductive biases, not guarantees. Grades combine knowledge, engagement, assessment practices, and circumstances. The defensible name is **course-performance readiness**, rather than a measured innate or causal cognitive ability.

**5. Research map: what is established and what is transferable**

The review uses primary papers, proceedings, author manuscripts, and the original MIRT monograph record. It is a targeted review rather than an exhaustive systematic review. Recent preprints are labeled as such. The complete source catalog records reading depth, publication status, URLs, and project-specific limits.

| Research line | Representative work | What we can use | What we should not import automatically |
| --- | --- | --- | --- |
| Psychometric person–item modeling | MIRT, Reckase 2009 [R02]; SPARFA 2014 [R04] | Course-adjusted performance factors; sparse low-dimensional structure; explicit scale restrictions | Named skills and comparable cross-faculty ability scores without anchors or connectivity evidence |
| Cognitive diagnosis | G-DINA 2011 [R03]; NeuralCD 2020 [R09] | Alternative interaction assumptions; monotonic student–requirement relationships | Expert Q-matrices, exercise text, or genuine concept labels absent from our contract |
| Noncompensatory requirements | Dealbreaker, ICML 2016 [R06] | A weak required dimension can limit success; smooth bottleneck functions | Claiming the bottleneck idea is new, or assuming all course grades follow an AND rule |
| Course-level knowledge accumulation | CKRM, SDM 2017 [R07]; nonlinear/attentive extensions, JEDM 2020 [R10] | The closest task precedent: prior course outcomes, provided/required knowledge, and concurrent-course context | Calling latent course requirements or target-conditioned history attention a new contribution |
| Temporal knowledge tracing | BKT 1994 [R01]; DKT 2015 [R05]; DKVMN 2017 [R08] | Predict before updating; maintain an evolving state | Pretending semester course grades form an ordered sequence of within-semester quiz responses |
| Difficulty-aware neural tracing | AKT, KDD 2020 [R12]; simpleKT, ICLR 2023 [R15] | Keep item identity/difficulty; use strong simple controls | Assuming elaborate attention or vector difficulty necessarily outperforms a simpler model |
| Fast probabilistic inference | VIBO, 2020 manuscript [R11]; VTIRT, EDM 2023 [R16] | Efficient approximate inference, local evidence, temporal uncertainty | Using future-conditioned trajectory smoothing to construct prospective training-query states |
| Sparse and new-student diagnosis | BETA-CD, AAAI 2023 [R14]; ICDM, WWW 2024 [R17] | Learned population priors; inductive states; uncertainty | Treating new-student inference as a solution to unseen-course requirements |
| Generative and changing-domain diagnosis | Generative CD, 2025 version [R18]; D-BIRD, 2025 preprint [R19]; MetaCD, 2025 preprint [R20] | Separate inference from prediction; consider cohort dynamics and changing support | Assuming known course calibration, future cohort labels, or inexpensive adaptation |
| Current 2026 directions | PAKT, July preprint [R21]; CLEAN, October preprint [R22] | Distinguish different learning mechanisms; examine stability of diagnostic representations | Repeated concept practice or verified expanding concept topology, which our data do not establish |
| Practical tabular learning and evaluation | TabM, ICLR 2025 [R23]; pyKT, NeurIPS 2022 [R13]; DomainBed, ICLR 2021 [R25] | Strong efficient controls, precise data boundaries, explicit model selection | A broad benchmark win or invariance penalty as proof of benefit on our historical holdout |

Several overlaps are decisive. CKRM already predicts next-term grades from grade-weighted provided knowledge and target requirements. NeuralCD already uses monotonic proficiency–difficulty interactions. Dealbreaker already uses a soft weakest-dimension likelihood. BETA-CD already learns priors with uncertainty, and ICDM/Generative CD already support inductive diagnosis. These facts substantially narrow our novelty claim.

VTIRT is especially useful for efficiency, but its full trajectory inference combines previous, current, and future ability potentials. That is legitimate retrospective inference. Our prospective query cannot use outcomes from its own semester or later semesters. Its workplace analysis also excludes learners with fewer than five interactions, so its reported results do not establish performance for our zero-history regime [R16].

D-BIRD's reviewed formulation separates cohort and individual trajectories while assuming known, pre-calibrated item difficulties. We would need to estimate course parameters under selected enrollments and changing cohorts, a harder problem. Its cohort component cannot be fitted on a future cohort's outcomes and then claimed as prospective robustness [R19].

PAKT's phase decomposition presumes concept-tagged, repeated practice. CLEAN's architectural invariance uses a known concept topology. Both are relevant to novelty checking, but neither is a direct drop-in design for these caches. A current paper's claimed causal interpretation or benchmark improvement is not evidence that its assumptions hold here [R21, R22].

The plausible research contribution is a carefully specified **semester-level, inductive, difficulty-aware readiness inference system for course failure prediction under uneven transcript and course support**, together with evidence that its structure improves ranking over strong existing models. Combining known ingredients can be valuable, but a publication-level novelty claim would require a more focused additional review and convincing empirical comparisons.

**6. The actual data contract restricts the design**

| Available or established | Design consequence |
| --- | --- |
| Global canonical student/course IDs, faculty labels, semester index, binary outcome | Build chronological student episodes and explicit identity/support audits |
| Numeric historical grades on 0–10 where present; grade below 5 agrees with failure | Use a consistent graded-outcome likelihood and binary marginal when grade is missing |
| Rich cached tabular features; loader uses the intersection of feature names across selected faculties | Preserve their predictive capacity, while auditing availability and avoiding accidental feature loss |
| Twelve course static fields plus five historical course statistics | Course encoder has 17 semantic fields; 34 values after preprocessing means values plus presence flags, not 34 independent descriptors |
| Course IDs and counts | Estimate supported course-specific parameters; shrink scarce ID effects and use an explicit unknown-ID fallback |
| Earlier-semester outcomes assumed released before the next semester | Freeze state before every query semester and update once outcomes are available |
| No grade-release timestamps | No claimed same-semester update order; test an additional semester embargo if timing is uncertain |
| Training frames preserve numeric grades, while target grades are discarded from model data | Training grade supervision is feasible through an adapter; evaluation target grades remain inaccessible to model inputs |
| Local joblib examples are synthetic smoke/validation caches | Do not invent new production statistics or empirical Suite28 results from them |
| Upstream cache generators and real production features are not fully available here | Audit/rebuild as-of course aggregates and document remaining raw-feature provenance before interpreting model gains |

The twelve static fields are `SoTinChi`, `SoTietMH`, `SoTietTuan`, `SiSoMax`, `SiSoMin`, `KhgTinhDTB`, `IsMHTN`, `IsMonKT`, `IsMHKN`, `IsMHDacThu`, `IsGDTiengAnh`, and `IsMonNgoaiNgu`. Historical fields are count, grade mean, grade standard deviation, failure rate, and pass rate.

The supplied contract does not establish verified concept tags, syllabi, prerequisite graphs, admissions assessments, attendance streams, withdrawal causes, instructor IDs, or problem-level dialogues. We should not design the initial experiment as though these inputs exist. Nonnumeric grade statuses are not accepted by the current grade validator, so a separate cause-of-failure or dropout head lacks a verified training target.

Two new courses with identical available descriptors and no outcomes may be indistinguishable to a metadata-only encoder. No sophisticated prior can recover their distinct content requirements without additional information. Course-ID effects help supported courses; they do not solve that fundamental unseen-course limitation.

Cross-faculty latent coordinates require care. The student–course observation graph may contain weakly connected or disconnected curriculum groups. For the initial design, use the same architecture with **faculty-scoped parameter/state blocks**, and compare predictions within the existing faculty evaluation cells. Do not report an Italian-language factor as psychometrically equivalent to an IT factor merely because both occupy coordinate 2.

Assemble histories using canonical global student IDs. Within each faculty block, calibrate any historical courses used for that faculty's student episodes from permitted pre-origin outcomes. Cross-faculty historical events may be used when their identity, outcome scale, and block-specific calibration are supported; otherwise route them through the declared metadata fallback and report that support separately. This preserves access to available transcripts without silently claiming cross-faculty scale equivalence. It is one framework with locally appropriate parameter blocks, not fourteen unrelated architectural designs.

**7. Three concrete Suite28 candidates**

| Candidate | Core predictor | Distinct commitment | Main advantage | Main risk | Relative implementation cost |
| --- | --- | --- | --- | --- | --- |
| A. Contextual Dynamic MIRT | Positive course discrimination applied to a causal uncertain readiness state, plus one scalar course difficulty | Largely compensatory course requirements | Cleanest statistical reference; economical; easiest to diagnose | A linear projection may miss course-specific bottlenecks; limited novelty | Low to moderate |
| B. Dynamic Readiness–Requirement Model (DRRM) | Smooth requirement deficits, integrated over a causal readiness distribution | Weak dimensions can matter disproportionately; readiness drives the score | Strongest fit to the intended multidimensional requirement hypothesis | Wrong compensation assumption, weak factor identification, or poor Gaussian approximation | Moderate |
| C. Joint Transcript–Tabular Model | Structured transcript state/gaps plus current features in a flexible neural tabular predictor | Predictive flexibility takes priority over a strictly psychometric decoder | Preserves nonlinear tabular interactions and can use state features selectively | Can bypass the structured theory and reproduce history overfit; weaker interpretability | Moderate |

These cost labels compare architecture work, not measured hours. All three require the same chronological data preparation and fair reporting. Candidate B is the recommended scientific lead. Candidate A is a compulsory control if B is chosen. Candidate C is a credible alternative if the user prioritizes flexible predictive capacity over the requirement theory, but it should not be quietly included to conceal a failing structured decoder.

**Candidate A: Contextual Dynamic MIRT**

Use a K-dimensional readiness distribution inferred from strictly earlier evidence, a positive discrimination vector `a_c`, and a single difficulty `b_c`:

$$
P(Y=1\mid\theta,c)=\sigma\left(b_c-a_c^\top\theta\right),\qquad a_{c,k}\geq 0.
$$

Here Y=1 is failure. A scalar difficulty is sufficient for the linear decoder; do not add a free requirement vector whose only effect is another course intercept. Normalize or regularize discrimination and anchor the latent scale. Use the same contextual semester forecast, graded observation likelihood, and support policy described for B. A K=1 version tests whether the data support anything beyond general readiness.

This is a strong reference architecture, not a claim of novel IRT. Under a Gaussian prior and an affine binary logistic likelihood, posterior inference is much better behaved than for a nonlinear bottleneck. Candidate A could beat B, and that would be a useful result: course failure might be mainly a compensatory or general-risk phenomenon.

**Candidate B: Dynamic Readiness–Requirement Model**

Maintain one readiness distribution per student, faculty block, and semester. A course decoder represents requirements as thresholds on the same latent axes. Historical grades update readiness through that decoder; the next course is predicted before its outcomes enter the state. Details follow in Sections 8–12.

The scientific upside is that transcript evidence is interpreted in the space of course demands rather than reduced to another weighted average of raw failures. The model also represents where the evidence is weak. The decoder makes a stronger structural commitment than MIRT, so it must earn that commitment empirically.

**Candidate C: Joint Transcript–Tabular Model**

Use the same chronological transcript representation, but feed its state mean, uncertainty, course-adjusted gaps, and support indicators together with current valid features to a compact MLP or parameter-efficient TabM-style predictor. Learn the final score directly from the outcome label; no prior-suite score, frozen tree residual, or posthoc tree correction is required [R23].

Prefer joint training to an in-sample supervised-feature stack. If a separately pretrained state generator is later used, create training representations by chronological cross-fitting and fit its transformations within each training origin. Otherwise course parameters or state features optimized on a row's own target could exaggerate the apparent downstream gain.

This model can outperform the structured predictor precisely because it is less constrained. That is its practical attraction and its scientific limitation. A win would validate useful transcript representations, not necessarily a faithful ability–requirement explanation. Compare against the identical tabular predictor without transcript features to identify whether the representation adds value.

**8. Recommended B design: one prediction boundary and five component roles**

For a target semester t, let `H_<t` contain only released earlier outcomes, `x_st` denote current student/semester covariates available at the prediction time, and `m_ct` denote legitimate course descriptors and earlier course statistics. Faculty subscripts are omitted from the equations; states and calibration operate within the declared faculty block.

| Component | Owns | Excludes | Evidence that it earns its role |
| --- | --- | --- | --- |
| Contextual state forecast | Current profile, valid academic summaries, elapsed time, observed enrollment burden where available | Query outcome and invented post-enrollment information | Useful no-history predictions; incremental value of filtered history over the same contextual model |
| Course requirement encoder | Supported course identity, metadata, earlier outcome calibration | A second unconstrained difficulty intercept or invented concept labels | Requirement diversity; stability; supported/rare/unseen-course comparisons |
| Outcome likelihood | Grade probabilities and their failure marginal | Contradictory independently trained grade/failure heads | Grade-bin consistency, failure ranking, calibration, grade-missing comparisons |
| Chronological evidence update | Released outcomes interpreted through the course likelihood | Same-semester outcomes before scoring; normalized attention averaging of raw failures | Correct sign of updates; support/direction-sensitive covariance; useful subsequent predictions |
| Temporal learner and evaluator | Next-semester loss, origin-local preprocessing, bounded selection, checkpointing | Model selection on the target holdout or future feature normalization | Later-origin AP against strong baselines; reproducibility and measured runtime |

There is no theorem that these components are “perfectly complementary.” This division removes obvious role conflicts and makes the complementarity claim observable.

**8.1 The state is shared across target courses**

Represent readiness before semester t by a Gaussian approximation:

$$
q^-_{s,t}(\theta)=\mathcal{N}\left(\mu^-_{s,t},\Sigma^-_{s,t}\right).
$$

Start with K=4 and a full 4×4 covariance. The full covariance matters: a historical course constrains a direction, not necessarily one independent axis. A diagonal approximation may falsely regard several correlated factors as separately measured. K=1 and a compensatory K=4 predictor remain useful controls; K=8 is a later, bounded check if there is evidence that additional dimensions are supported.

There is no student-ID embedding optimized separately for every person. Identity connects a permitted transcript to an episode; it does not supply a memorized risk parameter. Reuse the same pre-semester state for every current target course in that faculty block. The target course changes the decoder, not the student's inferred readiness vector.

Do not give the raw target-course row wholesale to the state encoder. Some columns describe the target rather than the student. Before implementation, assign every retained field to student/semester context, course/query context, or unavailable information. Student/semester fields must be consistent across a student's simultaneous query rows. Course-specific retake and prior-course information requires an explicit course/query representation; it cannot silently make the purported common student state target dependent.

The initial requirement decoder can use verified course/query covariates, but preserve a declared intrinsic course threshold separately from enrollment-specific burden. If a valid row-specific burden is included, express it as a small, centered effective-hurdle adjustment, shared across requirement dimensions, with no course-only intercept hiding inside it. Exclude unclassified fields explicitly and report the excluded feature list. A full-feature tabular control tests whether this structured partition loses important signal.

**8.2 Rich features inform the forecast, not a second independent transcript likelihood**

Let `u(x_st)` be a compact neural map from verified student/semester features to current expected readiness. Forecast from the previous posterior using:

$$
\mu^-_{s,t}=\rho_\Delta\mu^+_{s,t-1}+(1-\rho_\Delta)u(x_{s,t}).
$$

$$
\Sigma^-_{s,t}=\rho_\Delta^2\Sigma^+_{s,t-1}+(1-\rho_\Delta^2)V(x_{s,t})+\Delta Q.
$$

Use a shared, bounded retention coefficient within a faculty block, depending on elapsed semesters; `V` is positive definite and `Q` is a positive diagonal process-noise matrix. Initialize a student without history from `u(x)` and `V(x)`. This is an explicitly conditional state transition, not a claim that GPA summaries and the grades from which they were computed are independent observations.

Academic summaries can be useful current covariates even when the supplied transcript is unavailable. They should not be multiplied into a second evidence likelihood as though they were fresh measurements. This matters for zero-history rows: missing raw transcript support and missing all academic information are different regimes.

The transition adds process uncertainty and weakens the old mean's influence as the gap grows. Its total variance is not guaranteed to increase monotonically for every previous covariance, because it also returns toward the contextual variance. Report that behavior; do not claim an unproved uncertainty guarantee.

If retention tends to zero, the model may rely almost entirely on current features. That is a possible fitted outcome, not necessarily an optimization bug. It would weaken the claim that course-adjusted history adds value, and should be visible in the comparison with the contextual-only control.

**8.3 One course requirement representation**

For supported courses, use a metadata-derived threshold plus a regularized ID residual:

$$
r_{c,t}=h_\psi(m_{c,t})+s(n_{c,<t})e_c,\qquad s(n)=\frac{n}{n+\lambda}.
$$

Counts and metadata aggregates exposed to a query are strictly prior-semester quantities. Unknown IDs receive no learned residual. Fit course-ID parameters only from training outcomes available before the outer origin. Training an end-to-end model on permitted training labels is ordinary supervised fitting; the ban is on target-origin labels and future-derived query inputs.

The support shrinkage is a practical regularizer, not an exact Bayesian posterior for course parameters. In the minimal design, course parameters are shrinkage point estimates. The readiness covariance therefore does **not** represent all parameter uncertainty, grading drift, or unseen-course uncertainty. Log those support conditions separately. A variational course-parameter extension is a later option if sparse-course uncertainty demonstrably limits performance.

No verified Q-matrix exists. Let requirement thresholds themselves encode which axes materially constrain a course. Dimensions with sufficiently low thresholds contribute little to its bottleneck. Do not also learn an unconstrained relevance vector for the same soft bottleneck. The following expression illustrates why:

$$
\tau\log\sum_k w_{c,k}\exp\left(\frac{r_{c,k}-\theta_k}{\tau}\right)=\tau\log\sum_k\exp\left(\frac{r_{c,k}+\tau\log w_{c,k}-\theta_k}{\tau}\right).
$$

Only the effective thresholds are identified. A free weight and threshold can substitute for each other. The initial B decoder consequently uses fixed equal weights and one total requirement vector. It also omits a separate free course difficulty intercept. These choices remove two avoidable redundancies; they do not prove complete factor identification.

**8.4 A correctly signed, nonlinear requirement score**

Define the smooth shortage:

$$
D_{s,c,t}(\theta)=\tau\log\left[\frac{1}{K}\sum_{k=1}^K\exp\left(\frac{r_{c,t,k}-\theta_k}{\tau}\right)\right].
$$

Use positive global/faculty scale `a` if needed, and define success utility `u_sc=-a D_sc`. For the smallest pilot fix the scale and τ rather than making every course's temperature and discrimination free. Smaller τ emphasizes the largest shortage; larger τ moves toward averaging. The normalized soft maximum need not be nonnegative: readiness above the thresholds gives negative D and low failure risk. It is a signed deficit score, not an always-positive penalty.

With the failure threshold fixed at utility zero:

$$
P(Y=1\mid\theta,c,t)=\sigma\left(aD_{s,c,t}(\theta)\right).
$$

For fixed course parameters, scale, and context, increasing any readiness coordinate cannot increase failure probability. The gradient of success utility is positive in every coordinate. This is a genuine decoder property. It does not guarantee that a raw feature change has a monotonic effect, that readiness factors correspond to named skills, or that the inferred state is causally correct.

For illustration, two readiness vectors `(1,1)` and `(3,-1)` have the same mean. A purely averaged readiness rule treats them similarly for an equal-demand two-axis course. A bottleneck rule can treat the second student as more vulnerable because one required dimension is weak. That is useful only if the course's grading/outcome process actually has such a bottleneck. The example is geometry, not a forecast of our AP.

This decoder belongs to the established noncompensatory family represented by Dealbreaker [R06]. Our proposed contribution is its course-level causal inference/training setting and empirical utility, not invention of the soft-min/soft-max function. Candidate A is a structural comparator using a different course-response function; it is not claimed to be an exact nested special case of B's equal-weight decoder.

**9. One coherent grade/failure likelihood**

Retain more information than pass/fail where numeric grades are available. A practical initial likelihood uses grade bins `[0,5)`, `[5,6)`, `[6,7)`, `[7,8)`, `[8,9)`, and `[9,10]`. Ordered utility thresholds define a cumulative-logistic distribution:

$$
P(G<b_j\mid\theta,c,t)=\sigma\left(\kappa_j-u_{s,c,t}(\theta)\right).
$$

Fix `κ_5=0`; higher thresholds are increasing positive increments. Bin probabilities are differences of successive cumulative probabilities, with outside endpoints 0 and 1. The first bin is exactly the failure probability. Start with faculty-scoped ordered thresholds and shrinkage rather than freely fitting every course's entire grading scale.

For a known numeric grade, use its categorical likelihood once. If grade is missing but the binary label exists, use the binary marginal: first-bin probability for failure, and the sum of passing-bin probabilities for success. Do not multiply an independent failure likelihood by a grade likelihood that already determines the same failure label. That would double count one outcome and generally distort inferred confidence.

This sacrifices some within-bin grade precision in return for a stable, consistent observation model. A continuous bounded grade distribution is possible later, but it needs checks for rounding, grade heaping, and unusual zeros. A separate grade MSE head plus a binary head is easier to write but lacks the desired likelihood consistency.

The grade likelihood is supervised only with available training outcomes. Every training query is predicted from the state before that semester's outcomes are assimilated. Its grade can be a training target, never an input to its own forecast. Missing grades remain missing, with explicit indicators; zero is not an imputed missing grade.

There is an important assumption here: the ordinal link makes within-pass grades additional evidence about the same utility that controls failure. If passing-grade variation mostly measures a different process, that supervision can hurt AP. A binary-only matched control tests that assumption. Other suites already use grade information in various forms; adding grades is not itself proof of a new information advantage.

**10. Difficulty-aware updates with bounded computation**

After the semester's outcomes are assumed released, update using the unordered set of eligible outcomes `E_st`. The target distribution under a conditional-independence approximation is:

$$
\widetilde p^+_{s,t}(\theta)\ \propto\ q^-_{s,t}(\theta)\prod_{i\in E_{s,t}}p(o_i\mid\theta,c_i).
$$

The likelihood is logistic/ordinal and, for B, nonlinear in readiness. This posterior is not conjugate Gaussian. We therefore call the implementation an **approximate Gaussian evidence filter**, not exact Bayesian inference.

A practical first update is one damped Fisher-scoring step. At the prior mean, compute each observation's log-likelihood gradient `g_i` and expected Fisher matrix `F_i`. Use:

$$
\Lambda^+_{s,t}=(\Sigma^-_{s,t})^{-1}+\gamma\sum_{i\in E_{s,t}}F_i.
$$

$$
\mu^+_{s,t}=\mu^-_{s,t}+\delta(\Lambda^+_{s,t})^{-1}\gamma\sum_{i\in E_{s,t}}g_i,\qquad\Sigma^+_{s,t}=(\Lambda^+_{s,t})^{-1}.
$$

The expected Fisher is positive semidefinite; use stable Cholesky solves rather than explicitly constructing numerical matrix inverses. With a scalar utility, each event's Fisher has the form `I_u v_i v_i^T`, where `v_i` is the utility gradient with respect to readiness. At K=4 this is inexpensive. A one-step update is an approximation, not a converged Laplace fit or a guarantee of correct posterior covariance.

For the compensatory binary model, an unexpected failure pushes readiness downward along the course's measured direction. Passing a difficult course pushes it upward. Expected outcomes produce smaller changes. Information depends on the likelihood and direction, not a learned attention weight normalized to one across the transcript. Courses probing similar directions leave orthogonal directions less identified.

Sum a semester's gradients and information before the single update. Shuffling its row order must not change the result, apart from insignificant floating-point differences. If a second scoring iteration is needed, recompute against the same original prior and the same likelihood set; do not multiply the same outcomes into the evidence a second time.

Use `γ=1` for the ordinary likelihood approximation. A fixed discount such as `γ=0.5` is a possible predeclared control if within-semester dependence produces overconfidence. It is then a tempered/generalized update, not the exact posterior of the independent observation model. Damping, bounded updates, and a covariance floor are numerical safeguards to be logged, not concealed learned relevance mechanisms.

The greatest statistical risk for B is **ambiguous failures**. A failure may mean that any one of several dimensions is weak. A nonlinear bottleneck can produce a multimodal posterior, which one Gaussian represents poorly. Full covariance is better than a diagonal approximation but does not solve multimodality. If B's uncertainty and subsequent predictions fail while A works, do not claim that a larger latent vector will fix the issue. A small mixture or richer filter would be a new, costed research decision.

The semester likelihood also treats course outcomes as conditionally independent after accounting for readiness. Shared assessment practices, illness, and engagement shocks can violate that assumption. Check residual within-student/semester dependence and confidence by course count. A dedicated semester shock is a possible later extension; adding it immediately introduces another separation problem between readiness and transient effects.

Course completion is also a learning experience, not only a measurement. The initial filter estimates readiness from outcomes; it does not identify the causal knowledge gain supplied by each course. Its contextual transition can forecast changing readiness, but a predictable pass need not create a large evidence update even if the student learned substantially. If later-course errors systematically depend on which earlier courses were completed after conditioning on the inferred state, a small course-conditioned transition is worth a separate experiment. That would be a predictive transition, not an identified causal teaching effect. Do not add free “knowledge provided” vectors immediately and claim they are distinguishable from course requirements and student selection.

**11. Prediction integrates readiness uncertainty**

The primary output is:

$$
\widehat p_{s,c,t}=\mathbb{E}_{\theta\sim q^-_{s,t}}\left[P(Y=1\mid\theta,c,t)\right].
$$

Use a small fixed antithetic or quasi-Monte Carlo integration rule, shared consistently across evaluations. A proposed starting budget is 8 draws for training and 32 for evaluation; verify stability on development predictions before freezing it. This is an engineering proposal, not a measured optimum. Deterministic integration draws prevent a different random ranking every time a checkpoint is evaluated.

Uncertainty is not a rule saying “uncertain students fail.” In an affine logistic model, additional uncertainty often moves a probability toward 0.5; the direction depends on the mean, and nonlinear decoders are more complicated. Integrate the model distribution rather than adding an arbitrary positive uncertainty surcharge to a risk score.

Likewise, student covariance is not a full predictive confidence certificate. The minimal model does not integrate course-parameter uncertainty or uncertainty in the raw-feature pipeline. Show support and empirical calibration beside the score, without claiming that a narrow latent interval proves correctness.

The model changes risk ordering through course requirements, different inferred readiness states, and relevant uncertainty. A uniform intercept correction cannot account for the intended AP gain.

**12. Training, identification, and temporal generalization**

Train on chronological student episodes. For every training semester: forecast the shared state, predict all eligible courses, score the pre-outcome grade/binary likelihood, then assimilate that semester to predict later semesters. Course and context parameters are shared across episodes and optimized end to end. The final objective is predictive prequential negative log-likelihood plus declared parameter regularization; it is not the claim that we optimize an exact joint Bayesian ELBO.

For each observed course outcome, score the integrated predictive marginal: `-log(mean_draw p(outcome | theta_draw, course))`, using draws from that semester's pre-outcome readiness distribution. Averaging the conditional negative log-likelihood instead is a different objective and should not be silently substituted. Sum these row-level predictive log scores and declare the regularization separately; this is not a claim to maximize an exact joint marginal likelihood for the whole semester. Grade-present and grade-missing rows use the categorical and binary observations defined in Section 9.

The gradient flows through the approximate update to improve subsequent prediction. With analytic utility/ordinal derivatives, the small Fisher update is differentiable without general-purpose MCMC or a large inner optimization loop. Truncated backpropagation is a memory-budget option only after profiling; carrying a state through a detached boundary preserves its forward history while shortening its learning gradient. Report that distinction if used.

Use natural prevalence for the basic likelihood. If batches oversample rare failures, weight their contributions to recover the intended population objective. Choose checkpoints and the bounded architecture recipe using recent-origin macro AP. Proper likelihood fitting can help representation learning and calibration, but is not mathematically guaranteed to maximize AP. A within-cell ranking loss is a possible later ablation, not another initial default module; many pairwise losses primarily target AUC rather than AP.

For scale and identification, fix the failure utility threshold, initial latent units, and the pilot's decoder scale/temperature. Anchor contextual coordinates using a fixed training reference population, or explicitly regularize their location and spread. Shrink ID residuals and requirement magnitudes. Never normalize using target-origin outcomes or target-derived state statistics.

These restrictions remove obvious location/scale freedoms and discourage unstable solutions. They do not prove that individual dimensions are fully identifiable. Permutations, weak measurement directions, and disconnected curriculum groups remain possible. Do not force a covariance decorrelation penalty merely to obtain attractive “independent skills” plots.

Measure course requirement diversity after removing each course's common threshold level. If all centered requirement vectors are almost identical or their variation has effective rank one, the claimed multidimensional course structure is unsupported. Examine readiness covariance eigenvalues, update directions, and stability across seeds after permissible factor alignment. Without external anchors, label dimensions as latent factors, not “math ability,” “language ability,” or other invented traits.

Enrollment is selected rather than randomized. Course difficulty and readiness are estimated from observed students taking observed courses. Repeated enrollments and shared courses can improve connectedness, but no observational decoder proves a causal effect of ability, workload, or a prerequisite. This framework predicts risk; it does not establish that increasing a latent coordinate would cause an intervention to succeed.

Do not add a free future-semester embedding fitted from target outcomes. Known elapsed time, student stage, and enrollment context can inform the forecast. A later course-drift model must forecast from earlier course outcomes, with uncertainty; it cannot access the current course's observed failure rate. An invariance/adversarial loss also deserves a matched control rather than being presented as automatic protection against shift [R25].

**13. Which extra techniques are worth considering, and when**

| Technique | Decision for the first experiment | Reason |
| --- | --- | --- |
| Course-adjusted outcome inference | Core | Directly tests the new representation hypothesis |
| Rich contextual state forecast | Core, with an explicit feature-ownership audit | Preserves useful current information and supports missing transcripts |
| Small full-covariance readiness state | Core | Represents information directions without an enormous state |
| Consistent grade/binary likelihood | Core plus a binary-only control | Uses grade structure without contradictory or duplicated outcomes |
| Course-ID shrinkage and unknown-ID fallback | Core | Supported identities contain information unavailable in the generic descriptors |
| Semester set update and causal masking | Core | Matches the actual timing resolution |
| History-availability and course-ID dropout | Limited training robustness checks | Can prepare for incomplete support, provided covariates are masked consistently with the simulated regime |
| Chronological support/query training | Core through prequential prediction | Aligns state learning with subsequent outcomes without an expensive MAML loop |
| Variational course uncertainty | Deferred | Potential sparse-course benefit, but additional approximation and loss accounting |
| Explicit course/semester difficulty drift | Deferred | May improve course-relative ranking; requires proof that earlier data predict the drift |
| Mixture posterior or a separate semester shock | Deferred behind diagnostic evidence | Addresses ambiguity/dependence but adds model complexity and identification issues |
| Course-completion-conditioned state transition | Deferred behind later-course residual evidence | Outcomes measure readiness and courses may change it; the present data do not identify a causal teaching effect |
| Syllabus embeddings or expert Q-matrix | Data-dependent future branch | Could materially improve new-course demands; these inputs are currently unverified |
| Prerequisite graph, large Transformer, generic retrieval memory | Excluded from the initial core | No supplied structure or experiment evidence justifies their cost here |
| Latent dropout/withdrawal causes or separate innate ability | Excluded without additional labels/measurements | Not identifiable from our present binary/grade contract |
| Contrastive invariance, distillation from Suite25, residual correction, learned blending | Separate later experiments | They change the scientific question and could obscure whether the new structure works |

Transcript dropout should distinguish incomplete transcript access from a genuinely new student with no prior academic record. In the first case, valid cached GPA summaries can remain. In the second, those summaries must be absent too. Masking history while retaining impossible summaries would produce a misleading cold-start test. Similarly, course cold-start tests must hide identity-derived statistics as well as the ID residual when those statistics would not exist for a new course.

These changes do not create new evidence. They train the model to operate when some evidence is unavailable and must earn their value on later-origin or deliberately simulated support conditions.

**14. A bounded experiment that can falsify the proposal**

The next implementation should follow the selected architecture, not launch a large tournament of all three designs. If B is selected, the initial pilot has B, matched contextual-only B, and A's compensatory decoder. Strong existing Suite22 and Suite25 blend predictions remain compulsory aligned references; a fresh strong tabular refit is needed if the feature pipeline is materially changed.

**Stage 0: validate the prediction boundary and measure cost.** Inspect actual production field ownership, IDs, duplicate events, and transcript/course support. Rebuild the five historical course aggregates strictly by earlier semester rather than blindly trusting covered cache fields. Document any raw cached feature whose generator cannot be verified. Check train/target alignment and the grade threshold. Run a small representative faculty and a large faculty through one epoch to measure throughput, peak memory, and projection of the complete job. This is a runtime gate, not a promised hours estimate.

**Stage 1: two recent development origins, one declared seed and recipe.** Use the latest two eligible development origins, normally F4/F5 under our current historical convention, rather than choosing based on early origins where MCGB looked strongest. Report all faculties, natural zero-history support, genuinely unseen student IDs, and scarce/unseen course IDs. Start at K=4 with one decoder temperature and one update iteration. Permit only one predeclared second temperature if the first fit exposes an identifiable compensation problem. Do not run a large per-faculty hyperparameter search.

The matched contextual-only B disables filtered outcome assimilation during both training and evaluation while retaining legitimate current summaries, course inputs, and the same likelihood. This isolates additional value from the course-adjusted evidence path. Postfit masks and permutations are useful supplementary diagnostics, but label them explicitly; they do not replace retrained controls.

**Stage 2: only continue after a useful signal.** Add a K=1 comparison, a matched binary-only variant, and a second seed for the successful structure. Check whether requirements have more than one effective direction and whether uncertainty improves calibration or subsequent prediction. Expand only the dimension/likelihood choice that the evidence motivates.

**Stage 3: freeze and evaluate the historical holdout once.** Fit under each required pre-origin boundary, then evaluate the frozen recipe on F6/F7 and every faculty using aligned rows and the established macro definition. Do not tune a temperature, blend, evidence discount, or checkpoint on those target labels. If a later unseen semester exists, evaluate it separately as the confirmatory test.

A single shared architecture recipe across faculty blocks keeps the comparison understandable. A genuinely unsupported or failed faculty block must remain visible; silently dropping it would inflate macro AP. Record how metrics treat cells with no positives and report coverage before ranking models.

As a **proposed ambition**, an absolute AP improvement of roughly 0.02 over blend would place the historical two-fold macro near 0.39. This is a decision threshold for a meaningful result, not a forecast or a mathematically guaranteed attainable level. A tiny positive difference above 0.369526 would not meet the requested substantial-upgrade objective.

| Observation | Interpretation and next action |
| --- | --- |
| B beats context-only, A, and the strongest baselines across recent origins, with useful holdout transfer | The structured hypothesis earns further work; quantify uncertainty and seed stability |
| A beats B | Compensatory/general readiness is better supported; adopt A without insisting on bottlenecks |
| B matches context-only | Filtered transcript evidence does not earn its complexity; inspect whether the state is bypassed or the information is already summarized |
| B improves grade likelihood or calibration but loses AP | Useful distribution modeling is insufficient for the ranking objective |
| Gain appears only in one faculty or one older origin | No broad temporal breakthrough; examine support rather than averaging away the regression |
| Requirements or states collapse to one factor | Multidimensional claims fail; compare the simpler K=1 model |
| B fails on observed-history rows as Suite27 did | Course adjustment/temporal representation has not resolved the failure; do not blame only newcomers |
| B fails on unseen courses with generic metadata | A data/content limitation may dominate; seek richer course descriptors before expanding the network |
| Posterior precision grows while prediction worsens | Likelihood dependence, multimodality, grading mismatch, or approximate inference may be miscalibrated |
| No stable recent-origin improvement after the bounded pilot | Stop this path or choose C explicitly; do not automatically add another collection of modules |

A paired student-cluster bootstrap is appropriate for the existing macro comparison, preserving the same student's rows jointly across cells. Report exact support, confidence intervals, and seed variability separately. A bootstrap conditioned on a fitted model does not quantify uncertainty about another future semester.

**15. Required results and mechanism logging**

The eventual Suite28 package should print concise progress and preserve a detailed machine-readable record. Every run should include a source/config/data manifest, availability assumptions, phase timing, checkpoint state, and resume compatibility checks. Save per-row predictions with canonical identifiers so old and new models can be compared without relying on rounded console tables.

| Logging area | Required output | Question it answers |
| --- | --- | --- |
| Fold and faculty performance | AP, AUC, Brier/log loss, calibration, prevalence, row/positive counts, warning-budget recall/precision/F2 | Does the model improve the actual target and operational ranking? |
| Development versus holdout | Exact eligible-cell macro, per-fold values, strong reference deltas, paired support | Does early success transfer, and is the comparison complete? |
| Transcript availability | Zero/1–4/5–16/17+ eligible events, semesters observed, elapsed gap, missing grade rates | Is the result explained by available evidence? |
| Student identity support | Seen/unseen ID and available-history status separately | Are newcomers and missing transcripts being confused? |
| Course support | Known/rare/unseen IDs, prior count, metadata availability, fallback usage, historical cross-faculty calibration support | Does course inference transfer beyond supported IDs? |
| State geometry | Mean/spread, covariance eigenvalues, retention, update size, projected target-direction variance | Is inference meaningful or collapsing? |
| Requirement geometry | Centered-vector diversity/effective rank, ID residual size, support-stratified drift across training seeds/origins | Are different courses actually using different dimensions? |
| Likelihood and filtering | Grade-bin probabilities, failure-marginal consistency, update signs, Fisher conditioning, damping/floor counts | Is the claimed mechanism numerically and semantically valid? |
| Uncertainty | Predictive calibration by history/support/precision, integration stability on development rows | Is confidence useful rather than merely small? |
| Controls and diagnostics | Retrained contextual/compensatory/binary controls labeled separately from postfit masks/shuffles | Which structural assumptions earn their cost? |
| Runtime and recovery | Rows/episodes per second, phase hours, peak CPU/GPU memory, useful versus discarded work, origin/config hashes | Can the experiment finish within an acceptable budget and resume correctly? |

Only report interpretable factor names if independent course-content evidence justifies them. A pretty heat map of learned factors is insufficient. Example student traces should show permissible earlier events, their predicted likelihoods, updates, and the resulting subsequent forecasts without exposing personal identifiers in public artifacts.

**16. Why this should cost less, and what remains unknown**

The Suite27 audit attributes approximately 70.39% of the 48.204 unique measured phase-hours to selection, with substantial discarded work and repeated forest growth. Those numbers are phase sums, not the total elapsed week. The new design removes the custom NumPy tree grower and the reader/tree regrowth cycle entirely.

Prepare chronological episodes once per origin. During training, process a student's semester state once and reuse it for all target courses. Batch event likelihoods and 4×4 solves on the accelerator where available. Use compressed episode arrays, explicit masks, and immutable origin manifests. Recompute trainable states for the current weights; caching a previous epoch's states without declaring it would change the learning algorithm. Frozen evaluation states can be cached by checkpoint hash.

For N course events and E student-semester episodes, small-state Fisher accumulation costs roughly `O(N K²)` plus batched `O(E K³)` solves, excluding neural encoder work. It avoids a factor for every tree node and another for repeated forest regrowth. Full-history filtering also removes the need to repeatedly materialize a separate 32-event window for every target row. It does not make training cost independent of the number of students, semesters, or origins.

One origin/config fit per faculty block, a bounded pilot, and atomic checkpoints should substantially reduce avoidable work. GPU batching, episode lengths, raw-feature width, cross-faculty transcript scope, and data loading still determine actual throughput. The present workspace contains no genuine production-data benchmark for Suite28, so an “hours rather than days” guarantee would be invented.

**17. Decision to make before implementation**

Choose **B: Dynamic Readiness–Requirement Model**, with **A as its required structural control** and a contextual-only comparison. Its primary score should come from the readiness distribution and course requirements. Keep the initial model small, chronological, and explicit about its approximate inference. Use C only as a separately chosen alternative if flexibility becomes the priority or the structured hypothesis fails the pilot.

The decisive hypothesis is: **difficulty-adjusted transcript inference produces a more useful state for the next course than raw history summaries and current features already provide.** The second hypothesis is: **some courses require multidimensional bottlenecks rather than a single compensatory projection.** They must be tested separately.

This direction has a defensible reason to be tried. It has no demonstrated probability of a breakthrough yet. If the bounded recent-origin experiment does not show substantial, stable value against Suite22/Suite25 blend, the correct conclusion is to stop or change the modeling commitment, not to declare the design theoretically perfect and continue expanding it.

**Primary references and evidence boundaries**

**[R01] Albert T. Corbett; John R. Anderson (1994). Knowledge tracing: Modeling the acquisition of procedural knowledge.** Journal article, User Modeling and User-Adapted Interaction 4:253–278. [Primary source](https://link.springer.com/article/10.1007/BF01099821). Reading boundary: Publisher abstract and original bibliographic record.

**[R02] Mark D. Reckase (2009). Multidimensional Item Response Theory.** Springer monograph. [Primary source](https://link.springer.com/book/10.1007/978-0-387-89976-3). Reading boundary: Publisher bibliographic record and chapter structure, including coordinate systems and linking.

**[R03] Jimmy de la Torre (2011). The Generalized DINA Model Framework.** Journal article, Psychometrika 76:179–199; original issue April 2011. [Primary source](https://www.cambridge.org/core/journals/psychometrika/article/abs/generalized-dina-model-framework/93B2EF3A71833F12E8D8CFD9FEC23287). Reading boundary: Publisher abstract and original issue metadata; publisher migration date distinguished from original publication.

**[R04] Andrew S. Lan; Andrew E. Waters; Christoph Studer; Richard G. Baraniuk (2014). Sparse Factor Analysis for Learning and Content Analytics.** Journal article, JMLR 15:1959–2008. [Primary source](https://jmlr.org/papers/v15/lan14a.html). Reading boundary: Primary journal abstract.

**[R05] Chris Piech et al. (2015). Deep Knowledge Tracing.** NeurIPS 2015 proceedings. [Primary source](https://proceedings.neurips.cc/paper_files/paper/2015/hash/bac9162b47c56fc8a4d2a519803d51b3-Abstract.html). Reading boundary: Primary proceedings abstract.

**[R06] Andrew Lan; Tom Goldstein; Richard Baraniuk; Christoph Studer (2016). Dealbreaker: A Nonlinear Latent Variable Model for Educational Data.** ICML 2016, PMLR 48:266–275. [Primary source](https://proceedings.mlr.press/v48/lan16.html). [Reviewed full text](https://proceedings.mlr.press/v48/lan16.pdf). Reading boundary: Primary proceedings record and full-text model, smooth approximation, and identifiability discussion.

**[R07] Sara Morsy; George Karypis (2017). Cumulative Knowledge-based Regression Models for Next-term Grade Prediction.** SDM 2017 proceedings, pp. 552–560. [Primary source](https://epubs.siam.org/doi/10.1137/1.9781611974973.62). Reading boundary: Primary proceedings abstract; explicit model also verified in the authors' 2020 paper (R10).

**[R08] Jiani Zhang; Xingjian Shi; Irwin King; Dit-Yan Yeung (2017). Dynamic Key-Value Memory Networks for Knowledge Tracing.** WWW 2017 paper; arXiv preprint first posted 2016. [Primary source](https://arxiv.org/abs/1611.08108). Reading boundary: Author full text, read–predict–write order and concept-state interpretation.

**[R09] Fei Wang et al. (2020). Neural Cognitive Diagnosis for Intelligent Education Systems.** AAAI 2020, 34(4):6153–6161. [Primary source](https://ojs.aaai.org/index.php/AAAI/article/view/6080/5936). Reading boundary: Primary full-text factor definitions, Q-matrix inputs, and positive-weight monotonicity construction.

**[R10] Sara Morsy; George Karypis (2020). Context-aware Non-linear and Neural Attentive Knowledge-based Models for Grade Prediction.** JEDM 2020 journal article; author arXiv version reviewed. [Primary source](https://jedm.educationaldatamining.org/index.php/JEDM/article/view/428). [Reviewed full text](https://arxiv.org/html/2003.05063v1). Reading boundary: Author full-text CKRM equations, attentive/max pooling, concurrent-course modeling, and experimental setting.

**[R11] Mike Wu; Richard L. Davis; Benjamin W. Domingue; Chris Piech; Noah Goodman (2020). Variational Item Response Theory: Fast, Accurate, and Expressive.** 2020 arXiv version reviewed; publication status of this version not assumed. [Primary source](https://arxiv.org/abs/2002.00276). Reading boundary: Author full-text inference formulation, amortization, and ablation discussion.

**[R12] Aritra Ghosh; Neil Heffernan; Andrew S. Lan (2020). Context-Aware Attentive Knowledge Tracing.** KDD 2020 paper. [Primary source](https://arxiv.org/abs/2007.12324). Reading boundary: Author full text, Rasch-based difficulty embeddings and multi-concept evaluation caveat.

**[R13] Zitao Liu et al. (2022). pyKT: A Python Library to Benchmark Deep Learning based Knowledge Tracing Models.** NeurIPS 2022 Datasets and Benchmarks. [Primary source](https://proceedings.neurips.cc/paper_files/paper/2022/hash/75ca2b23d9794f02a92449af65a57556-Abstract-Datasets_and_Benchmarks.html). Reading boundary: Primary proceedings abstract.

**[R14] Haoyang Bi et al. (2023). BETA-CD: A Bayesian Meta-Learned Cognitive Diagnosis Framework for Personalized Learning.** AAAI 2023, 37(4):5018–5026. [Primary source](https://ojs.aaai.org/index.php/AAAI/article/view/25629). [Reviewed full text](https://ojs.aaai.org/index.php/AAAI/article/download/25629/25401). Reading boundary: Primary full-text hierarchy, support/query meta-objective, and variational adaptation.

**[R15] Zitao Liu; Qiongqiong Liu; Jiahao Chen; Shuyan Huang; Weiqi Luo (2023). simpleKT: A Simple But Tough-to-Beat Baseline for Knowledge Tracing.** ICLR 2023. [Primary source](https://arxiv.org/abs/2302.06881). Reading boundary: Author full text and difficulty ablations; acceptance confirmed by arXiv record.

**[R16] Yunsung Kim; Sreechan Sankaranarayanan; Chris Piech; Candace Thille (2023). Variational Temporal IRT: Fast, Accurate, and Explainable Inference of Dynamic Learner Proficiency.** EDM 2023 short paper. [Primary source](https://educationaldatamining.org/EDM2023/proceedings/2023.EDM-short-papers.24/index.html). Reading boundary: Primary full-text Gaussian potentials, forward/backward aggregation, and workplace data exclusions.

**[R17] Shuo Liu; Junhao Shen; Hong Qian; Aimin Zhou (2024). Inductive Cognitive Diagnosis for Fast Student Learning in Web-Based Online Intelligent Education Systems.** The Web Conference 2024, pp. 4260–4271. [Primary source](https://arxiv.org/abs/2404.11290). Reading boundary: Author full text, inductive split, and explicit limitation on new exercises.

**[R18] Jiatong Li; Qi Liu; Mengxiao Zhu (2025). Generative Cognitive Diagnosis.** July 2025 arXiv v1 reviewed; later publication status not assumed. [Primary source](https://arxiv.org/html/2507.09831v1). Reading boundary: Author full text, generative diagnosis/response separation and expert-Q inputs.

**[R19] Hansol Lee; Jason B. Cho; David S. Matteson; Benjamin W. Domingue (2025). Dynamic Bayesian Item Response Model with Decomposition (D-BIRD): Modeling Cohort and Individual Learning Over Time.** June 2025 preprint; record says submitted to AIME-Con. [Primary source](https://arxiv.org/html/2506.21723v1). Reading boundary: Author full-text cohort/individual decomposition, known-item-difficulty assumption, and sampling approach.

**[R20] Jin Wu; Chanjin Zheng (2025). MetaCD: A Meta Learning Framework for Cognitive Diagnosis based on Continual Learning.** December 2025 arXiv preprint. [Primary source](https://arxiv.org/abs/2512.22904). Reading boundary: Author abstract and submission record.

**[R21] Duantengchuan Li; Yingqian Bi; Jinsong Chen; Rui Zhang; Mingwen Tong (2026). Disentangling Knowledge States with Ability and Proficiency Modeling for Knowledge Tracing.** July 2026 arXiv preprint version reviewed. [Primary source](https://arxiv.org/html/2607.13103v1). Reading boundary: Author full-text phase decomposition and multi-branch Transformer; treated as a preprint.

**[R22] Tao He; Jinxing Xiang; Fan Jiang (2026). CLEAN: Psychometrically Consistent Incremental Cognitive Diagnosis under Concept-Space Expansion via Architectural Isolation.** October 1, 2026 arXiv preprint; no completed venue inferred from template text. [Primary source](https://arxiv.org/html/2610.02278v1). Reading boundary: Author full-text abstract, invariance mechanism, and expert-Q topology requirement.

**[R23] Yury Gorishniy; Akim Kotelnikov; Artem Babenko (2025). TabM: Advancing Tabular Deep Learning with Parameter-Efficient Ensembling.** ICLR 2025; arXiv v3 February 2025. [Primary source](https://arxiv.org/html/2410.24210v3). Reading boundary: Author full-text abstract and introduction, plus primary publication record.

**[R24] Zachary Lipton; Yu-Xiang Wang; Alexander Smola (2018). Detecting and Correcting for Label Shift with Black Box Predictors.** ICML 2018, PMLR 80:3122–3130. [Primary source](https://proceedings.mlr.press/v80/lipton18a.html). Reading boundary: Primary proceedings abstract and label-shift assumption.

**[R25] Ishaan Gulrajani; David Lopez-Paz (2021). In Search of Lost Domain Generalization.** ICLR 2021 paper; author full-text arXiv version reviewed. [Primary source](https://arxiv.org/abs/2007.01434). Reading boundary: Author full-text model-selection argument and controlled ERM comparison.

Local empirical sources are the completed Suite27 log/review and archived Suite26 source packages. The actual production dataset and upstream cache generators were not re-executed for this proposal. All Suite28 equations, engineering budgets, expected benefits, and acceptance targets are proposed design choices or analytical arguments, not new experimental findings.
