Three research gaps motivated this thesis. The proposal set out to close Gaps 1–2 by building an embedding-free, cold-start, auditable personalization system. Pursuing the auditability requirement to its conclusion produced the primary scientific novelty: a governance-observability instrument (the ADC) and a calibrated causal topology finding.
Thesis Identity: Cold-Start ITS — auditable & reproducible by constructionAnchor: N=96,727 · run-d2154070 · seal-bae44d1aCausal seal: N≈1.98M · p<0.01
§1
Section 1.2
Problem Formulation — Three Gaps
1Embedding Dependency & Interpretability
Current methods (DKT, SAKT, GKT) depend on latent learner representations that require extensive interaction data, are difficult to interpret, and introduce computational overhead. During cold-start — where only limited interaction data are available — these representations are unreliable or degenerate (zero-padded LSTM state, random GNN embedding).
2Cold-Start Personalization
Personalization quality in early interactions depends not just on how much data exists, but on the quality and coverage of interaction signals. Insufficient exploration leads to limited interaction support — the system keeps recommending what it already knows, producing unreliable mastery updates and misleading personalization.
3Governance ObservabilityPRIMARY GAP
Adaptive systems using multi-dimensional governance functions — weighting mastery, recency, transfer, efficiency, engagement, and challenge simultaneously — have no existing instrument to characterize which dimensions carry signal under observed conditions. All evaluations measure prediction accuracy only. This leaves two systems with identical AUC scores potentially having dramatically different internal governance structures, invisible to current methods.
§2
Section 1.2 cont.
Three Contributions + Dependency Flow
Contribution Dependency
AEmbedding-Free Auditable Runtime — the realized proposal, PRIMARY deliverable
↓ its auditability requirement, pursued to its conclusion, yields
The system the proposal set out to build. Effective cold-start personalization without per-learner embeddings, every decision traceable and auditable.
Architecture: Event-sourced HCIE with outbox-first state mutation and Kafka-mediated typed consumers. Append-only log guarantees deterministic replay.
Mastery Ensemble: Two-learner Kalman + Bayesian fusion (V2 deployed). No offline training. Third "Lyapunov" learner runs in V1 sealed but cut from canonical fusion (r=0.92 redundant with Bayesian).
Bandit: Phased Thompson-sampling over modalities (text/MCQ/video/audio/code). Beta(α,β) posterior per (learner, concept, modality). Real learners only.
JT Attribution: Six-dimensional Joint-Objective score recorded per interaction to deterministically replayable trajectory store. Enables post-hoc ADC audit.
Success Evidence
Cold-start from first interaction without embeddings. Deterministic replay proven. Explicit per-interaction JT attribution. End-to-end governance lineage: Kafka event → DAG lookup → JT → mastery prediction.
B
PRIMARY SCIENTIFIC NOVELTY
ADC + Calibrated Topology Finding
The proposal's auditability requirement, pursued to its logical conclusion: an instrument that characterizes which governance dimensions carry signal, and a calibrated causal finding.
ADC Definition: Sealed thresholds α_floor=0.01, signal_ratio>0.08. Classifies each governance dimension as ACTIVE or structural_zero from the empirical JT attribution distribution. Never modifies the runtime.
Topology Taxonomy: Predicts governance activation from dataset structure across five topology classes, eight datasets.
Reflexive Calibration: ADC audited its own headline verdict — found a normalizer floor artifact (σ(−2.5)≈0.076 non-zero floor even on null dimensions). Categorical "ACTIVE" was threshold-fragile. Re-derived under shuffled-DAG control.
Causal Finding: Shuffled-DAG control (same graph density/degree, randomized edges) separates correct topology from graph presence. Durable causal component ≈+0.053, p<0.01, N≈1.98M. ~⅔ is curriculum proximity; time-placebo removes ≈42% learner selection.
Validates that the embedding-free runtime of Contribution A is practically sound — where ADC confirms the transfer dimension is ACTIVE.
Protocol: Matched baseline: same 10 held-out learners, same interaction windows, same cold-start protocol. No information advantage for HCIE.
Mandatory floor: Beat BKT = required. Sealed V1 (3-learner, Junyi): HCIE 0.609 vs BKT 0.612 — within 0.003, effectively tied. Deployed V2 (2-learner+prior) beats BKT on ASSISTments at every window.
Cold-start win: HCIE beats BKT by +0.037 to +0.044 AUC at ≤5 and ≤10 attempts. Mechanism: KF uncertainty + ZPD drives exploration before BKT EM converges on sparse data.
Honest caveats: DKT/SAKT lead on high-base-rate dense corpora overall. Junyi BKT is degenerate at cold-start (≤5 AUC=0.51). Two pooled per-window losses are Simpson's-paradox artifacts.
Success Evidence
HCIE beats BKT on every cold-start window (ASSISTments-2009). Beats BKT, DKT, SAKT on CSEDM at every window. BKT floor holds on all 4 datasets. Competitiveness conditional on ADC-ACTIVE topology, not a universal superiority claim.
HCIE 0.6051 leads all baselines on pooled overall AUC (matched-eval m_K, tie-aware). HCIE−BKT margin: +0.0088 (n=10) → +0.0062 (n=40) → +0.0125 (n=76, the only statistically significant point). The earlier graph-eval 2-learner companion readout (HCIE 0.609 / BKT 0.612, a different rejected config) is shown only for transparency and is NOT the headline.
Only ≤5 favours HCIE; ≤10/≤20 favour BKT, and every per-window CI crosses zero — no per-window claim is robust. The robust result is the pooled overall lead above (HCIE +0.0125, significant at n=76).
§4
Section 1.3
Formal Thesis Claim
Primary Thesis Claim (Governance Instrument). This thesis introduces the Adaptive Dimension Controller (ADC) — an observational instrument that characterizes which dimensions of a multi-dimensional governance function carry empirical signal under an observed interaction ecology. The ADC never reads back into the runtime: it does not gate, mask, or reweight any dimension; it only classifies the recorded attribution distribution.
Central empirical demonstration — prerequisite topology. Under an ability-matched shuffled-DAG control, explicit prerequisite topology contributes a real, causal, cross-concept, dose-dependent transfer effect. Shuffled-DAG ≈ 0, p < 0.01. The majority is attributable to curriculum proximity (same-session sequencing). Durable causal component ≈ +0.053 within-learner scale. Time-placebo removes ≈42% learner selection. Full-corpus seal: N ≈ 1.98M, p < 0.01.
Formal properties of the instrument:
Operates at inference time without modifying the runtime or requiring offline training
Classifies governance dimensions from empirical JT attribution distribution (α_floor=0.01, signal_ratio > 0.08)
Reports structural_zero about its own headline when the normalizer floor artifact was found
Separates correct topology from mere graph presence via shuffled-DAG control — a control the graph-based KT literature generally omits
Validated by reflexive calibration: the topology effect is the instrument's honest demonstration, not its definition
§5
Section 1.3.1
Scientific Maturation — From Categorical to Calibrated
A governance-observability instrument earns trust only if it is willing to report a structural_zero about its own headline. The ADC's first sealed verdict was stress-tested in three stages:
1
Internal Audit — Normalizer Floor Artifact Found
The headline's normalizer was a centered sigmoid with no zero-guard, producing a non-zero floor (σ(−2.5) ≈ 0.076) even on genuinely null dimensions — proven to five decimals on a live record. The categorical "ACTIVE" verdict was therefore threshold-fragile: classified mean contributions sat near the normalizer floor, not well above it.
→ structural_zero filed against own headline — instrument is honest
2
External Control — Shuffled-DAG Separates Topology from Graph Presence
An ability-matched shuffled-DAG control (same graph density and degree distribution, randomized edges) was run. A committed, conclusion-blind estimator on raw Junyi outcomes re-estimated the effect under within-learner fixed effects, experience/recency controls, same-family vs cross-family split, the shuffled-DAG null, and a time-placebo negative control.
→ shuffled-DAG ≈ 0; observed effect survives — topology is doing real work, not just graph presence
3
Calibrated Finding — Effect Frozen to What is Validated
Explicit prerequisite topology contributes a real, causal, dose-dependent transfer effect. Cross-concept (not confined to same-skill-family). About two-thirds is curriculum proximity. A time-placebo attributes ≈42% of the durable association to learner selection, leaving a durable causal component ≈ +0.053. Full-corpus seal: N ≈ 1.98M, p < 0.01. This supersedes an earlier ≈+0.020 figure whose computation was not retained.
Maturation pattern (documented twice for this thesis):Freeze the claim to what is validated. First instance: froze unvalidated pedagogical-effectiveness claims to the validated event-sourced runtime. Second instance (here): froze the categorical topology claim to the controlled empirical effect. The instrument that performed the calibration is the contribution; the calibrated effect is its honest demonstration.