Nexus
Web & Desktop
The additive cognitive diagnostic models arrive. A-CDM, LLM and R-RUM join G-DINA, DINA and DINO in one engine — six models, one Q-matrix, compared with AIC and BIC — in calibration, in the Skill Profiler, and from Claude Desktop.
What's new
A-CDM, LLM and R-RUM
The three classic additive models (de la Torre, 2011): G-DINA with the interaction terms removed, so each required skill adds a fixed main effect on the probability scale (A-CDM, identity link), the log-odds scale (LLM, the linear logistic model, logit link) or the log scale (R-RUM, log link) — k + 1 parameters per item instead of 2k. They run on the same multi-start EM as G-DINA with a Newton (Fisher-scoring) M-step and were verified against the GDINA R package to the convergence tolerance. Every item still reports its pattern probabilities, δ, guess, slip, IDI and the monotonicity check.
A model chooser that explains the link
The Model option is now a chooser in three families — Saturated (G-DINA), Conjunctive / disjunctive (DINA, DINO) and Additive (A-CDM, LLM, R-RUM) — with a one-line description of the selected model. For G-DINA, DINA and DINO the Link Function remains a reporting scale for δ; for the additive models it is part of the model, so the selector locks and says fixed by the model. The item table shows the model and its link, and the fit cards count parameters per model so all six can be compared with BIC on the same data.
Enforce monotonicity
A new Estimation Options checkbox (off by default; the R package's
mono.constraint) constrains every skill main effect of an additive model to be non-negative, so P(correct) can never fall when a skill is mastered. The other models keep the monotonicity check. The item plot now charts P(correct) by attribute pattern for every model.Skill Profiler
Profiles examinees under all six models. The pattern-level parameter file gains a
model column (G-DINA when absent), and the profiler's model toggle — now two rows, guess/slip and pattern-level — locks to the loaded calibration or to the file's model. Unknown model names are reported instead of silently becoming DINA.Claude Desktop and the AI Advisor
calibrate_cdm accepts ACDM, LLM and RRUM plus mono_constraint, rejects a link that conflicts with an additive model, and writes the model column; score_cdm reads it back and locks to that model. The AI Advisor's knowledge base covers the additive models, the two roles of the link function, the Newton M-step and the monotonicity constraint.Learn it
The CDM Calibration and Skill Profiler tours describe the six models in all eight languages. PACER University's CDM module adds a step, Additive Models: When the Link Is the Model, with a six-model fit comparison. The technical reference documents the additive models, their design matrix, and the constrained Newton M-step.
Fixes
CDM API
An unknown CDM method or link name now returns a clear 400 error instead of silently calibrating or scoring with DINA and the identity link.
Web application
Windows desktop
macOS desktop