For reviewers — run it yourself

Watch the auditable loop run, step by step

This is one real sealed learner, from a brand-new start. Drag the slider (or press ◀ ▶) to step through their interactions and watch the closed loop the thesis is built on — observe → estimate → state → govern → decide. Click any stage for the why. The numbers are frozen and reproducible.

Interaction 1 / 140cold-start ≤5concept: ext_graph_representing_numbers

Observe

✓ correct

Estimate · 2 learners

Kalman m

Bayesian 0.42

Explicit state

0.503

±0.035 unc.

Govern · JT 6-dim

transfer

0.0156 ⚡

ZPD 0.93 · JT 0.56

Decide · Thompson

ext_junyi_graph_task_representing_numbers

external_log:junyi_2015_graph

1. Observe — why it matters (RM1) A brand-new learner answers one question. No profile, no embedding, no training data — just this correct/wrong signal on a concept. This is the cold-start condition the whole thesis is about.

The learner's mastery, building from nothing

Canonical mastery (m_K) per interaction. The dot is where you are — click anywhere on the line to jump. Starts uninformed, climbs as evidence accrues — visibly, auditably.

What this loop buys — the real contribution

The headline isn't the AUC (that's competitive — cold-start is hard for everyone). It's the structural axes below that embedding models cannot do at all:

capability	HCIE	BKT	DKT	SAKT	GKT
Cold-start estimate, no training	✓	✓	✗	✗	✗
Embedding-free	Full	Limited	✗	✗	✗
Update cost / interaction	O(1)	O(1)	O(d²)	O(M²d)	O(L·\|E\|·d)
Per-interaction governance (JT 6-dim)	✓	✗	✗	✗	✗
Decision traceability & audit	✓	✗	✗	✗	✗
Governance transparency (ADC)	✓	✗	✗	✗	✗
Sealed reproducibility	✓	✗	✗	✗	✗
Representation personalization	✓	✗	✗	✗	✗
Cold-start predictive AUC	Competitive	Competitive	Strong*	Strong*	Weakest

*Strong on dense data; at true cold-start all models sit at AUC ≈ 0.57–0.64 — see the per-window evidence. Thesis Tabel 4.15.

Verify each claim yourself

The numbers (RM5)

Cold-start AUC, every model & dataset, run picker.

Topology → transfer (RM4)

Shuffled-DAG control: real effect at p<0.001.

Governance (RM3)

Which JT/ADC dimensions carry signal.

Sealed & reproducible (RM3)

Fingerprint → re-run → identical.

The live tutor (RM1·RM2)

Drive the real loop on a fresh account.

Reproducible: ex_junyi_graph_239464 (sealed, anonymized) · run-94a3b8ba / seal d2154070 · content-hash 85690d8b…. 140 interactions, frozen — re-run the seal and the fingerprint is identical.