For each theoretical framework: what it claims, which paper introduced it, how HCIE applies it, which component implements it, and whether the implementation is faithful (Tier 2 grounding status).
DKT h_t ∈ R^d requires offline training; a cold h_0 = zeros is degenerate. SAKT needs sequence history. GKT node states are global.
Per-(learner,concept) state (μ,σ²) + (α,β) — O(2) space, O(1) update, working from the first interaction via the population prior.
Kalman filter + Beta-Bernoulli conjugate pair
Gap 2 — Cold-Start Exploration
BKT/DKT are purely reactive — they update on what is observed, with no explicit exploration. Insufficient diversity leads to unreliable mastery updates.
The Kalman variance σ² feeds the JT Uncertainty signal, so the system recommends under-explored concepts. Thompson Sampling over modalities adds explicit exploration.
All prior methods are evaluated by AUC alone. Two systems with identical AUC may have completely different internal governance structures.
The ADC classifies each governance dimension as ACTIVE or structural_zero from the sealed JT attribution distribution. Post-hoc, with zero blast radius.
ADC (novel) + preregistered threshold design + shuffled-DAG causal control
§2
Click ▼ for the formula + CP notes
Framework Cards
Math/StatsPASS
Kalman Filter (Scalar)
Kalman (1960). "A New Approach to Linear Filtering and Prediction Problems." Welch & Bishop (2006) for scalar formulation.
What It Claims / Provides
Optimal linear-Gaussian state estimator: given a linear state-space model with Gaussian noise, the Kalman filter computes the minimum mean-squared-error posterior over the hidden state. Prediction + update in O(1) per observation.
How HCIE Uses It
One of two canonical mastery estimators. Tracks per-(learner, concept) ability mean μ and variance σ². The variance term σ² feeds the Uncertainty JT signal — high σ² = unexplored concept = the system recommends it.
Component: kalman_learner.py
Tier 2 audit: Sound mathematical framework. Scalar Kalman update equations verified correct. Predictive validity r=0.3322 — best single learner in the ensemble.
Math/StatsPASS
Bayesian Beta-Bernoulli (Conjugate Pair)
Corbett & Anderson (1995). "Knowledge Tracing: Modeling the Acquisition of Procedural Knowledge." J. of User Modeling and User-Adapted Interaction.
What It Claims / Provides
The Beta distribution is the conjugate prior for the Bernoulli likelihood: if the prior is Beta(α,β), then after observing a Bernoulli outcome r the posterior is Beta(α+r, β+1−r). Posterior mean = α/(α+β) = the Maximum A Posteriori mastery estimate.
How HCIE Uses It
Second canonical mastery estimator. A per-(learner, concept) Beta(α,β) posterior. Starts from a population prior (Yudelson et al. 2013 individualized-prior). Simple, interpretable, and immune to cold-start.
Component: bayesian_learner.py
Tier 2 audit: Sound conjugate-Bayesian framework. O(1) update. Correlation with outcomes PASS. Works from the first interaction via the informative prior.
ML/OnlinePASS
Thompson Sampling (Bayesian MAB)
Thompson (1933). "On the Likelihood that One Unknown Probability Exceeds Another." Agrawal & Goyal (2012). "Analysis of Thompson Sampling for the Multi-armed Bandit Problem." Lattimore & Szepesvári (2020). "Bandit Algorithms."
What It Claims / Provides
Thompson Sampling achieves Bayes-optimal expected regret O(√(KT log T)) for K-armed Bernoulli bandits over T rounds. Posterior sampling naturally balances exploration (high variance) and exploitation (high mean) without a separate exploration parameter.
How HCIE Uses It
Selects the learning modality (text/MCQ/video/audio/code) per (learner, concept) pair. Each arm maintains a Beta(α,β) posterior over its success probability. An arm is chosen by sampling θ_a ~ Beta(α_a, β_a) and serving argmax θ_a.
Component: bandit.py
Tier 2 audit: Thompson Sampling is mathematically sound. The live MAB loop is proven in the DB (112k+ trajectory rows, 801 users). Guard: real learners only — synthetic traffic is excluded from posterior updates.
Math/StatsDISCLOSE
Item Response Theory — 1PL (Rasch Model)
Rasch (1960). "Probabilistic Models for Some Intelligence and Attainment Tests." Embretson & Reise (2000). "Item Response Theory for Psychologists."
What It Claims / Provides
P(correct | ability θ, difficulty b) = σ(θ − b), where σ is the logistic. Item difficulty b and learner ability θ live on the same scale; their difference determines the probability of success. Separability theorem: item and person parameters are identifiable in the 1PL model.
How HCIE Uses It
Referenced for the Challenge/BaselineDifficulty signal. V1 uses inverted latency as a proxy for difficulty — explicitly disclosed as NOT a faithful IRT implementation. A true IRT 1PL (estimating b per item) is deferred to V2.
Component: unified_brain.py (Challenge signal)
Tier 2 audit: IRT is the intended framework, but the V1 implementation (inverted-latency proxy) does not faithfully implement it. Explicitly disclosed in §4.5 and renamed BaselineDifficulty by locked design decision. An IRT 1PL rewrite is planned V2 work.
PedagogyDISCLOSE
Zone of Proximal Development (ZPD)
Vygotsky (1978). "Mind in Society: The Development of Higher Psychological Processes." Harvard University Press.
What It Claims / Provides
The ZPD is the range between what a learner can do independently and what they can do with guidance. Effective instruction targets tasks slightly above current mastery — neither too easy (boring) nor too hard (frustrating).
How HCIE Uses It
JT signal 5: ZPD(i,k) = 1 − |ability(i) − difficulty(k)| / range. It peaks when ability and difficulty are matched, pushing the system to recommend "just-above" concepts rather than ones that are too easy or too hard.
Component: unified_brain.py (JT signal 5)
Tier 2 audit: The ZPD signal is mathematically sound but saturates at ~0.97 across the sealed run — limited discriminative signal. This means the system scores nearly every concept equally on ZPD. Disclosed in §4.5 and the JUSTIFICATION_LEDGER.
ControlDISCLOSE
Lyapunov Stability (BoundedStability heuristic)
Lyapunov (1892). "The General Problem of the Stability of Motion." Referenced as stability motivation; not a formal Lyapunov theorem in the implementation.
What It Claims / Provides
A dynamical system ẋ = f(x) is stable if there exists a Lyapunov function V(x) > 0 with dV/dt ≤ 0. It provides a certificate of bounded, non-divergent state trajectories without solving the full ODE.
How HCIE Uses It
Third mastery estimator in the V1 ensemble (BoundedStability). It applies a contraction-style update to prevent mastery oscillation — but it is NOT a formal Lyapunov function. Cut from the V2 canonical fusion (r=0.92 correlation with the Bayesian learner = redundant).
Component: lyapunov_learner.py (V1 only; weight=0 in V2)
Tier 2 audit: Disclosed as a bounded heuristic, not a formal Lyapunov stability guarantee. It runs in V1 sealed and is recorded in the trajectory store, but is excluded from the canonical ensemble fusion. The synergy audit found r=0.92 with the Bayesian learner — near-redundant, cut per the §4.8.2 decision.
ML/OnlineDISCLOSE
Exponentiated Gradient Ensemble (EG)
Kivinen & Warmuth (1997). "Exponentiated Gradient versus Gradient Descent for Linear Predictors." Littlestone & Warmuth (1994). "The Weighted Majority Algorithm."
What It Claims / Provides
EG maintains weights w_l over L experts, updated as w_l ← w_l · exp(−η · loss_l) / Z. Regret bound: Σ_t loss_best + O(log L / η) total excess loss over T rounds — logarithmic in L, sub-linear in T.
How HCIE Uses It
Updates the ensemble weights across the Kalman and Bayesian learners. Weights adapt based on each learner's prediction loss per interaction. In V1 sealed, weights are near-uniform (CV=0.03–0.06), meaning EG barely moved from its initialization.
Component: unified_brain.py (ensemble fusion)
Tier 2 audit: The EG update is mathematically correct. However, sealed V1 shows weight CV=0.03–0.06 — near-uniform, barely moved from init. The ensemble's predictive r=0.3113 < Kalman solo r=0.3322. Disclosed: the ensemble does not currently outperform its best component.
Math/StatsDISCLOSE
Prerequisite DAG Transfer (T_realized)
Barnes (2005). "The Q-matrix Method." Pavlik et al. (2009). "Performance Factors Analysis." Chen et al. (2018). "Prerequisite-Driven Deep Knowledge Tracing."
What It Claims / Provides
In a directed acyclic graph of prerequisites, mastery of predecessor concepts transfers positive signal to learning the successor concept. Formally: P(correct_k | mastered_prereqs) > P(correct_k | not_mastered_prereqs).
How HCIE Uses It
T_realized = Σ_{p ∈ prereqs(k)} mastery(i,p). It aggregates prerequisite mastery as a transfer signal for concept k and feeds JT signal 2. It requires explicit DAG edges — bipartite Q-matrices leave it dormant.
Component: unified_brain.py (JT signal 2)
Tier 2 audit: Target-blind: T_realized uses graph presence (an edge exists) rather than edge strength, so it does not distinguish strong from weak prerequisites. The causal evidence is the shuffled-DAG result: the real DAG yields +0.053 durable transfer, while a random DAG yields ≈ 0.
Novel contribution of this thesis. Sealed threshold design inspired by preregistered hypothesis testing conventions (Simmons et al. 2011; Nosek et al. 2018). Signal-ratio threshold analogous to coefficient of variation tests in psychometrics.
What It Claims / Provides
A governance dimension d is ACTIVE iff: (1) E[contribution_d] > α_floor, AND (2) std(contribution_d)/E[contribution_d] > signal_ratio_threshold. Otherwise it is structural_zero. The instrument runs post-hoc on sealed trajectory data and never modifies the runtime.
How HCIE Uses It
Characterizes which JT governance dimensions carry empirical signal under the observed interaction ecology. It classifies each of 6 dimensions as ACTIVE or structural_zero. This is the thesis's primary scientific contribution.
Tier 2 audit: The ADC is the instrument performing the calibration — and it also reported structural_zero on its OWN normalizer floor artifact (reflexive calibration). This self-audit is the primary evidence of the instrument's validity.
All state changes are represented as an immutable, append-only sequence of events. Current state = the replay of all events from origin. The Transactional Outbox pattern guarantees the atomic write of a domain event plus an outbox row in a single DB transaction, preventing lost messages.
How HCIE Uses It
The substrate that makes ADC measurement possible. All JT attribution, mastery updates, and recommendations are written atomically with outbox rows → Kafka → typed consumers → the trajectory store. This enables deterministic replay for any historical window.
Tier 2 audit: Mutation authority validated. Side-channel mutations are architecturally prohibited. Replay determinism is demonstrated (the same event stream + seed → the same JT scores). End-to-end governance lineage is proven: Kafka event → DAG lookup → JT → mastery.
BaselineN/A
BKT Baseline (Corbett-Anderson HMM)
Corbett & Anderson (1995). "Knowledge Tracing: Modeling the Acquisition of Procedural Knowledge." UMUAI 4(4). Yudelson et al. (2013). "Individualized Bayesian Knowledge Tracing Models" for individualized priors.
What It Claims / Provides
A two-state HMM: latent state K_t ∈ {0,1} (unlearned/learned). Four parameters: P(L₀), P(T), P(G), P(S). The EM algorithm finds the MLE of these parameters from response sequences. The individualized prior (Yudelson) estimates a separate P(L₀) per learner from the first N attempts.
How HCIE Uses It
The mandatory evaluation floor: HCIE must beat or match BKT. V1 sealed: 0.609 vs 0.612 — tied. V2 deployed (2-learner + individualized prior) beats BKT at every window on ASSISTments-2009.
Component: Evaluation baseline only
Tier 2 audit: An external baseline, not an HCIE component. Evaluation protocol: the same held-out learners and the same interaction windows. BKT is degenerate on Junyi cold-start (≤5 AUC=0.51 — near-random).
CausalPASS
Time-Placebo Negative Control
Novel to this thesis (§3.3.6, §4.6). Design principle from difference-in-differences causal inference. Related: Angrist & Pischke (2009). "Mostly Harmless Econometrics."
What It Claims / Provides
A negative control (placebo) is an exposure that cannot cause the outcome through the hypothesized mechanism. In the time-placebo design: if a learner will master prerequisite A *after* attempting target B, that ordering is causally impossible for A→B transfer — so any association is pure selection bias.
How HCIE Uses It
Decomposes the topology transfer effect into (a) a durable causal component and (b) learner selection. Treatment = "the learner crossed edge A→B before attempting B" (a real causal path). Placebo = "the learner will cross edge A→B after attempting B" (no causal path). The difference is the causal effect, which eliminates ≈42% of the learner-selection confound.
Tier 2 audit: Implemented in the committed estimator. b_treatment=0.099, b_placebo=0.041, durable causal ≈ +0.053. The placebo attributing 42% to selection is EXPECTED and strengthens the claim — it shows the instrument is detecting real confounds.
CausalPASS
Within-Learner Fixed Effects
Standard panel data / econometrics method. Wooldridge (2002). "Econometric Analysis of Cross Section and Panel Data." Hsiao (2014). "Analysis of Panel Data."
What It Claims / Provides
By demeaning observations within each unit (learner), fixed effects eliminate all time-invariant unobservable confounders specific to that unit. The causal effect is identified from within-unit variation only — effectively controlling for learner ability, motivation, and every stable characteristic.
How HCIE Uses It
Applied in the committed estimator for the topology transfer effect. It controls for learner-specific ability, motivation, and study habits — all of which correlate with both graph traversal (the treatment) and correct responses (the outcome).
Tier 2 audit: A standard econometric technique, applied correctly. Combined with the time-placebo, it isolates the topology-specific causal effect to a ≈ +0.053 durable component.
CausalPASS
Permutation / Randomization Test
Fisher (1935). "The Design of Experiments." Good (2000). "Permutation Tests." Applied here as shuffled-DAG permutation: K=100 random edge relabelings.
What It Claims / Provides
Under the sharp null hypothesis (no treatment effect), the observed test statistic is equally likely to be any permutation of the data. The p-value = the fraction of permuted statistics ≥ the observed one. It is an exact finite-sample test — no distributional assumptions.
How HCIE Uses It
Establishes the statistical significance of the shuffled-DAG topology effect. K=100 permutations of the edge labels (preserving the degree sequence) form the null distribution. p = #{null_k ≥ observed} / K < 0.01.
Tier 2 audit: p < 0.01 with K=100 permutations on a 1/10 sample; the full-corpus seal (N≈1.98M) confirms. The shuffled-DAG null is exactly the right null for topology — it destroys the prerequisite ordering while preserving the graph statistics.
PedagogyDISCLOSE
Cognitive Load Theory + Multimedia Learning
Sweller (2024). "Cognitive Load Theory and Educational Technology." Mayer (2020). "Cognitive Theory of Multimedia Learning." Ainsworth (2021). "How Multiple Representations Support Learning." Scheiter & Gerjets (2007).
What It Claims / Provides
CLT: working memory has limited capacity (7±2 chunks); learning fails when intrinsic + extraneous load exceed capacity. Multimedia learning: dual-channel processing (visual + auditory) allows more total throughput. Multiple representations: different formats (text/diagram/equation) activate different cognitive schemas.
How HCIE Uses It
The theoretical justification for the modality bandit's arm set (text/MCQ/video/audio/code). The hypothesis: some learners have a lower cognitive load with visual formats (video), others with symbolic ones (code/MCQ). The bandit discovers this per learner via Thompson Sampling rather than assuming one modality fits all.
Tier 2 audit: CLT/multimedia learning is the pedagogical motivation for the modality bandit, not a computational framework that was directly implemented or verified. The representation-selection arm was inactive on the V1 sealed path — it is only active in live deployment. Disclosed as designed-but-not-yet-sealed.
Math/StatsDISCLOSE
T_prospective — 5 Failed Formulations
Chen et al. (2018). "Prerequisite-Driven Deep Knowledge Tracing." Knowledge-Space Theory (Doignon & Falmagne 1999). DINA model (de la Torre 2009). PSI-KT (structural drift). All referenced in §3.3, p.118.
What It Claims / Provides
Prospective transfer: P(learn concept B | currently mastering A, where A→B in the DAG). In theory, the set of concepts reachable from the current mastery frontier determines learning potential, and Dijkstra reachability gives a natural scoring function.
How HCIE Uses It
Designed as JT signal 6 — Σ_{successors(k)} (1−mastery(i,successor)). All five tested formulations either (a) failed topology-specificity (the shuffled-DAG also showed an effect), (b) were redundant with Challenge/difficulty, or (c) introduced computation that exceeded the O(indeg(k)) bound. Held at 0.0 in V1. The ADC's ability to classify this as structural_zero validates the instrument.
Tier 2 audit: Explicitly dormant. Five formulations were tested: Knowledge-Space outer-fringe, DINA noisy-AND, PSI-KT structural drift, prerequisite-depth utility, and Dijkstra reachability score. All are disclosed in §3.3, p.118, and deferred to future work. The ADC correctly classifies it as structural_zero — this is a validation of the instrument, not a failure.
Math/StatsDISCLOSE
Population Prior — Individualized BKT (Yudelson)
Yudelson et al. (2013). "Individualized Bayesian Knowledge Tracing Models." ITS 2013. Enables per-learner initial mastery estimation from population-level base rates.
What It Claims / Provides
Standard BKT uses a single P(L₀) for all learners on a concept. Individualized BKT fits a separate P(L₀)_i per learner from their first N interactions — reducing cold-start AUC error through better initialization. The population distribution P(L₀) across learners serves as the prior for new learners.
How HCIE Uses It
The V2 deployed configuration uses the Yudelson individualized prior as the Beta(α₀, β₀) initialization for the Bayesian learner. The population mean/variance of correct responses per concept seeds α₀ = E[p]·κ, β₀ = (1−E[p])·κ, where κ is a concentration parameter. This is what pushes V2 to beat BKT at every cold-start window.
Tier 2 audit: The individualized prior is the V2 deployed configuration; V1 sealed used a uniform Beta(1,1). The prior computation is audit-only in the trajectory store (materialised via tier2_5_continuation_run.py) — it is not yet promoted to the causal compute_jt path. Future work per §6.3.
§3
What this thesis adds to the literature
Named Results & Contributions
Shuffled-DAG Control (Novel)
Separates correct prerequisite topology from mere graph presence. Null: a random DAG with the same degree distribution. Treatment: the real DAG. Permutation p < 0.01, K=100.
This control is absent from the GKT, GIKT, and DyGKT papers. They show an AUC improvement with graphs but do not separate topology from presence.
ADC Topology Taxonomy (Novel)
Predicts governance activation from dataset structure across 5 topology classes (explicit DAG / bipartite Q-matrix / flat skill tags / transition graph / null). The transfer dimension is ACTIVE iff an explicit prerequisite DAG is present.
8 datasets classified; transfer dormant on 4 of 8 (bipartite/flat/transition/null)
The first instrument to characterize the structured dormancy of governance dimensions rather than measuring AUC alone.
Reflexive Calibration Pattern (Documented)
The instrument audits its own headline: the normalizer floor artifact σ(−2.5)≈0.076 was found and reported. The categorical "ACTIVE" verdict was frozen to a controlled empirical estimate. The pattern is documented twice for this thesis.
α_floor=0.01 < normalizer_floor=0.076 → threshold-fragile → re-derived under shuffled-DAG
"Freeze the claim to what is validated." First: pedagogical-effectiveness was frozen to the event-sourced runtime. Second: the topology claim was frozen to the controlled effect.
Cold-Start AUC Floor (Empirical)
HCIE beats BKT at ≤5, ≤10, ≤20, and overall on ASSISTments-2009. Mechanism: Kalman uncertainty drives exploration before BKT's EM converges.
Zero offline training cost. The cold-start advantage shrinks as history grows (BKT's EM catches up at warm-start) — an expected trade-off, reported honestly.
§4
Adaptive Dimension Controller — the primary scientific contribution
ADC Deep Dive — CP Walkthrough
The ADC as a Read-Only Monitoring Pipeline
Sealed trajectory store
experiment_trajectories
N=96,727 rows (immutable)
→
Aggregate per dimension
μ_d = mean(jt_{d})
σ_d = std(jt_{d})
→
Two-threshold classify
μ_d > α_floor AND
σ_d/μ_d > ratio_thresh
→
Topology taxonomy
5 dataset classes →
activation prediction
→
Reflexive calibration
audit own verdict →
file structural_zero if needed
Invariant: The ADC reads only. There is no write path to the runtime. Blast radius = zero. Every number it produces traces back to a sealed trajectory row.
1. The Classification Algorithm
// ADC classification — runs post-seal on frozen trajectory snapshot// Input: experiment_trajectories WHERE run_id = '<sealed_run>'// Output: verdict[d] ∈ {ACTIVE, structural_zero} for each dα_floor = 0.01 // pre-registered, sealed before data analysisratio_threshold = 0.08 // pre-registeredDIMENSIONS = { "ΔM": "jt_delta_m_contribution", "T_realized": "jt_transfer_contribution", "Challenge": "jt_challenge_contribution", "Uncertainty": "jt_uncertainty_contribution", "ZPD": "jt_zpd_contribution", "T_prospective": "jt_transfer_prospective_contribution",}for name, col in DIMENSIONS: values = SELECT col FROM experiment_trajectories WHERE run_id = sealed_run μ_d = mean(values) σ_d = std(values) ratio = σ_d / μ_d // coefficient of variation // TWO-THRESHOLD GATE: if μ_d > α_floor AND ratio > ratio_threshold: verdict[name] = "ACTIVE" // dimension carries signal else: verdict[name] = "structural_zero" // dimension is dead weight// Sealed V1 results (run-94a3b8ba, N=96,727):// ΔM: μ=0.083 cv=0.43 → ACTIVE// Challenge: μ=0.158 cv=0.22 → ACTIVE ← largest contributor// Uncertainty: μ=0.047 cv=0.38 → ACTIVE// ZPD: μ=0.052 cv=0.15 → ACTIVE (near structural_zero border)// T_realized: μ=0.031 cv=0.29 → ACTIVE (on explicit-DAG datasets)// T_prospective:μ=0.000 cv=N/A → structural_zero (hardcoded 0.0)
2. The Floor Artifact — Why α_floor ≠ 0 (the critical bug)
Bug report: The JT normalizer is a centered sigmoid σ(x) = 1/(1+e^(-x)). At x=−2.5: σ(−2.5) ≈ 0.076. This means even a genuinely null dimension (true signal = 0) gets a contribution of ~0.076 after normalization — not zero. So α_floor = 0.01 < 0.076 would classify a null dimension as ACTIVE. The categorical verdict was threshold-fragile.
// The normalizer (V1 — no zero-guard):def normalize_jt(raw_jt): return sigmoid(raw_jt) // = 1 / (1 + exp(-raw_jt))// At raw_jt = 0 (null dimension):sigmoid(0) = 0.5 // NOT zero — midpoint of sigmoid// At raw_jt = -2.5 (weak signal near null):sigmoid(-2.5) ≈ 0.076 // floor artifact// Result: "null" dimensions get mean_contribution ≈ 0.076// α_floor = 0.01 < 0.076 → threshold would classify them as ACTIVE// But they're not — they're just sitting on the normalizer floor// Why the ADC caught this:// It computed mean_contribution for T_prospective = 0.000 (hardcoded 0.0)// → T_prospective bypasses the sigmoid (explicitly zeroed)// → other "active" dimensions might also be near the floor// → the ADC filed structural_zero against its own verdict// → re-derived under shuffled-DAG control for topology-specific confirmation// The fix (V2 plan — F4):def normalize_jt_v2(raw_jt, zero_guard=True): if zero_guard and raw_jt == 0: return 0.0 // explicit zero → no floor artifact return max(0, sigmoid(raw_jt) - sigmoid(0)) // center at true zero
3. Topology Taxonomy — Predicting Activation from Dataset Structure
The ADC predicts whether the transfer dimension (T_realized) will be ACTIVE or structural_zero based on the dataset's topology class — before the instrument is even run. This is the taxonomy's predictive power.
T_realized = a real transfer
signal; shuffled-DAG ≈ 0
Bipartite Q-matrix
Items × skills matrix
(no skill→skill edges)
ASSISTments-2015,
most IRT datasets
structural_zero
No prereq edges → T_realized
always sums to 0
Flat skill tags
Single-level skill labels
(no hierarchy)
EdNet (flat KC),
Khan Academy basic
structural_zero
The predecessor set is empty
→ T_realized = 0
Transition graph
Empirical co-occurrence
(no explicit prereqs)
pyKT auto-generated
graph from response data
structural_zero
Edges encode co-occurrence,
not prerequisite direction
Null graph
No concept relationships
(single concept or random)
Single-skill KT tasks,
shuffled-DAG null
structural_zero
No edges → T_realized = 0;
shuffled-DAG confirms
// Taxonomy as a decision function (O(1) lookup):def predict_transfer_activation(dataset): if dataset.has_explicit_prerequisite_DAG(): return "ACTIVE" // multi-level concept→concept edges elif dataset.is_bipartite_Q_matrix(): return "structural_zero" // item×skill, no skill→skill edges elif dataset.has_flat_skill_tags(): return "structural_zero" // single-level taxonomy, no hierarchy elif dataset.is_transition_graph(): return "structural_zero" // empirical co-occurrence ≠ prerequisite else: return "structural_zero" // null / single-concept// Why this matters for SYSTEM SELECTION (not just evaluation):// If you deploy HCIE on a dataset with flat skill tags:// T_realized = 0 for all learners → the JT is effectively 5-dimensional// Challenge dominates (μ=0.158, cv=0.22) → system degrades to difficulty-based selection// If you deploy on explicit DAG (Junyi):// T_realized active → topology drives transfer signal → causal benefit ≈ +0.053// The taxonomy tells you IN ADVANCE which regime you're in
4. Reflexive Calibration — The Instrument Audits Itself
// Reflexive calibration = running the ADC ON the ADC's own output// This is unusual — most instruments don't self-audit// Step 1: Initial sealed verdict (pre-calibration)initial_verdict = { "ΔM": "ACTIVE", // μ=0.083 > 0.01 ✓ "Challenge": "ACTIVE", // μ=0.158 > 0.01 ✓ (largest contributor) "Uncertainty": "ACTIVE", // μ=0.047 > 0.01 ✓ "ZPD": "ACTIVE", // μ=0.052 > 0.01 ✓ "T_realized": "ACTIVE", // μ=0.031 > 0.01 ✓ "T_prospective":"structural_zero", // μ=0.000 (hardcoded 0.0)}// Step 2: ADC audits the normalizernormalizer_floor = sigmoid(-2.5) // ≈ 0.076// Observation: some "ACTIVE" μ values (0.031, 0.047, 0.052) are NEAR the floor// T_realized μ=0.031 << normalizer_floor 0.076 → inconsistent!// How can a real signal have lower mean than the floor of a null dimension?// Answer: T_realized is genuinely near-dormant on the specific topology tested// Step 3: ADC files structural_zero about its own headlinecalibration_event = { "type": "REFLEXIVE_AUDIT", "finding":"normalizer_floor_artifact", "detail": "σ(-2.5) ≈ 0.076 → null dimensions score > α_floor=0.01 → threshold-fragile", "action": "re-derive topology effect under shuffled-DAG control",}// Step 4: Re-derive under shuffled-DAG + time-placebo// (see Time-Placebo card for the causal estimator)calibrated_result = { "b_durable": 0.099, "b_placebo": 0.041, "causal_delta": "+0.053", "null_shuffled": "≈ 0", "p_value": "< 0.01", "verdict": "REAL causal topology effect — not just floor artifact",}// Why this is the contribution (not the original categorical verdict):// Any instrument can produce verdicts.// Only an instrument willing to REVISE its own verdict earns trust.// The ADC found its own floor bug, re-derived, and got a STRONGER result.// That chain — find bug → rerun under control → confirm effect — is the science.
5. ADC as a Software Engineering Pattern
CI/CD Test Suite
Each governance dimension = one assertion. ACTIVE = the test passes. structural_zero = the test fails. Running it post-seal = an integration test against a production snapshot.
// assert mean_contribution > 0.01
// assert cv > 0.08
// → PASS or FAIL per dimension
Distributed Systems Health Check
The ADC = a read-only health endpoint for the governance layer. It checks whether each JT signal is "alive" (carrying signal) or "dead" (stuck at the floor) and returns a per-dimension status.
GET /governance/health
→ {"ΔM":"ACTIVE",
"T_prospective":"zero"}
Coverage Tool (Instrumented)
The trajectory store is like a coverage report: every JT signal that fires gets a row. The ADC reads it and asks "was this code path actually executed with real variation, or is it just dead code?"
The ADC defines a contract: "governance uses 6 dimensions." structural_zero = the implementation violates the contract (the dimension exists in code but does nothing in practice).
// contract: dim ∈ ACTIVE
// violation: dim ∈ structural_zero
// → disclose in Methods section
The key invariant: The ADC is a post-hoc observer, not a runtime controller. It reads the immutable trajectory log, classifies dimensions, and reports. It NEVER writes back. This means: (a) zero blast radius on the live system, (b) findings are reproducible from the sealed snapshot alone, and (c) the instrument can be re-run with different thresholds without touching the runtime. This is the operational realization of the thesis's original auditability requirement.