Topology Taxonomy β transfer-dimension class match 8 / 8
ADC sealed thresholds (Ξ±_floor = 0.01, signal_ratio_threshold = 0.08) were frozen before any run data was observed. The taxonomy covers five topology classes across eight dataset/phase pairs. Only explicit_dag topology with runtime graph injection is predicted ACTIVE; all others are predicted structural_zero β an ecology property, not a system failure.
β PartialTopology prediction is robust; the transfer signal it gates is not yet anchored
The transfer-dimension taxonomy holds at 8 / 8 β sealed thresholds before run data, no leakage. This is a separate metric from self-prediction taxonomy accuracy, which is L4 18/24 = 0.75 (Tabel 4.11). But the transfer signal those thresholds gate currently responds to graph presence, not edge correctness (see /review/methods and the shuffled-DAG control on /dashboard/audit). The taxonomy is correct; what activation means for prediction gets calibrated in the sealed re-run.
Flips after:
Shuffled-DAG control re-run with sealed Phase-2 (real vs permuted vs no-DAG)
β Loadingβ¦Per-dataset run_id and N shown in the table below.
Live prerequisite graph
The concept DAG the transfer mechanism runs on
Live from the sealed Junyi dataset graph. Hover a concept to trace its transfer links; node size is its connectivity. This is exactly the structure the shuffled-DAG test below randomizes.
Loading graphβ¦
0.01
Ξ±_floor
Mean T_realized must exceed this
0.08
signal_ratio_threshold
std/mean ratio must be below this
8 / 8
Transfer-dimension class match
All matched before any run data
Loading...
Key insight: GKT (graph-aware baseline) trained on the same Junyi DAG achieves only AUC 0.571 overall β the worst of all evaluated models. This confirms that graph topology alone does not guarantee improved prediction; governance observability requires characterizing which dimensions are active, not just adding edges.
Shuffled-DAG Causal Decomposition
Separating correct topology from mere graph presence β a methodological contribution generally omitted by graph-based KT literature.
Observed Cross-Concept Effect
b = +0.091
Mastering prereq BEFORE target
Shuffled Control
b β 0 (p < 0.01)
Same graph density, edges randomized
Time Placebo
b = +0.038
Mastering prereq AFTER target (selection bias)
Placebo-corrected durable residual β NOT clean causal
placebo ratio β 0.42 β ~42% of the apparent past effect is state/selection confoundperm-p < 0.001 (K=1000) (K=100) β robust to permutation
This is a placebo-corrected residual, not net/clean causal transfer: curriculum proximity (b = +0.134) and learner state remain partially confounded with the effect. Coefficients estimated jointly; the decomposition is illustrative, not an orthogonal variance partition.
Decomposition β coefficient b
Coefficient b per condition, on a common axis β the residual (right) is what survives both the shuffled-edge and time-placebo controls.
N = 1,976,020first-encounters Β· sealed 2026-06-03 Β· p < 0.01 (permutation test) Β· Source: prospective_probe_v3_full.json
R12 ablation note: R12 ablation was attempted but withdrawn (confounded β cross-run state accumulation, sign-flip in multi-seed). Shuffled-DAG is the primary causal evidence.