One toolchain across
"modeling → weighting → merging"

Large assessments involve hundreds of thousands to millions of test-takers, many items, and multiple dimensions. Traditional IRT software fails at this scale — memory explodes or estimation never finishes — and commercial packages are hard to customize. Research teams need an engine that is both rigorous and able to run big data on an ordinary workstation.

IRTC is a self-contained marginal maximum likelihood (MML) estimation engine with C++ (Rcpp / RcppArmadillo) internals, systematically engineered for scale:

• Model coverage: Rasch/1PL, PCM, RSM, 2PL, GPCM; unidimensional and between-item multidimensional; latent regression, multi-group analysis, case weights, and EAP ability estimation.
• Performance: multithreaded parallel E-step, block-wise streaming posterior computation (no full person × node posterior matrix), dimensional decomposition, analytic gradients with hybrid Newton M-step, automatic engine routing.
• Trustworthiness: optional node pruning is accompanied by stratified empirical error checks with a written report — the exact path is the default, and statistical targets are never silently changed.

• The streaming engine has completed million-sample, multidimensional GPCM estimation in prototype benchmarks, compressing memory needs to ordinary-workstation levels.
• Three engines (auto / grid / streaming) with automatic routing that records its reasoning — usable by default, controllable when needed.
• 26 automated test suites and a beginner-friendly manual; engineering quality on par with established open-source projects.

Surveys often need overall and multiple overlapping subgroup pass rates (by region, gender, urban/rural) to approach or exactly match given targets — without breaking the original sampling weight structure or population margins. Manual reweighting is neither rigorous nor reproducible.

• Formalizes the requirement as bounded convex quadratic programming (QP), solved at the population-unit × outcome aggregation level — rigorous and efficient on large samples.
• Two modes — soft (approach targets) and exact (match exactly) — both preserving the initial weight structure and overall margins. Methodologically grounded in Deville & Särndal (1992) calibration theory.
• Practical tooling: one-step interface, pre-solve data checks, target-table builders, effective sample size (ESS) and design-effect diagnostics, example data; powered by the high-performance osqp solver.

• Published on CRAN under the MIT license, with 13 automated test suites.
• Turns a real-world rate-calibration challenge into a reproducible, diagnosable standard workflow.

Tabulating graded results by population units (province × gender × urban/rural × age × education) often leaves cells with tiny or zero samples, destabilizing proportion estimates. Units must be merged so that provincial and national subgroup grade shares (e.g., A/B/C/D) stay within target intervals and every final unit has positive samples — a constraint-heavy combinatorial problem that is extremely error-prone by hand.

• Solves within-province merging as set-partitioning mixed-integer linear programming (MILP).
• Lexicographic objectives: minimize merged sample size, then population distance, outcome heterogeneity, weight distortion, and merge count — in that order.
• Honest by design: when constraints are infeasible, it reports so explicitly instead of silently compromising the targets.

• Deterministic, reproducible merging plans replace cell-by-cell judgment calls.
• Open-sourced under MIT with 16 automated test suites; CRAN submission upcoming.