Verification registry

Every run, and the gates it survived.

Nothing appears on this page until five checks pass. The stamps on each entry are those checks; the certificate lets you recompute every number from the data.

T1 · Reality

Model-implied success matches the observed rate within 10 points.

T2 · Bootstrap

On data the model generated itself, fit statistics read ≈ 1. A broken meter fails here.

T3 · Invariants

No impossible values ship: fit means near 1, no |Z| > 10 unjustified, no perfect correlations.

T4 · Ledger

Every filter named, every row accounted for. Prose and table cannot disagree.

T5 · Two calculators

Engine and an independently written reference agree to 1e-6, item by item.

Public-data calibrations

Novel numbers on real public datasets
CalibrationRev · WO-HEFF-04

Finding Percents

ASSISTments 2012–13 · CML Rasch · WLE persons

212items
0.735reliability
+0.49targeting
T1T2T3T4T5
b3e1a70a … f2dc9b9
CalibrationRev · WO-HEFF-04

Ordering Fractions

ASSISTments 2012–13 · CML Rasch · WLE persons

71items
1.32separation
+0.88targeting
T1T2T3T4T5
c8d6dfeb … 96117eb

Synthetic recoveries

Known effects, planted and found — the engine’s exam

Two kinds of entries, honestly labeled. Synthetic recoveries reproduce published effects from data calibrated to published parameters. Public-data calibrations produce new numbers from real datasets — and carry revision notes whenever a defect was caught, because catching them is the product.