One in-house workflow across
“cleaning → modeling → weighting”

Survey and assessment data often arrive in multiple formats and encodings, with out-of-range values, missing responses, straight-lining, timing anomalies, and form-specific item identifiers. Ad-hoc scripts bury correction logic in code, making it hard to review why a value changed or reproduce the full run.

• Versioned rules run range, set, and expression checks plus detectors for missingness, straight-lining, response time, trap items, and score anomalies.
• Detection and execution are separate: exclude, set-NA, recode, and flag actions must map explicitly to rules, while unhandled findings remain visible.
• Raw data stay immutable; every change records old and new values, rule, method, and finding_id, while manifests and hashes support verified reruns.
• Common tabular and statistical formats, chunked processing, scoring, multi-form item-bank alignment, and self-contained HTML reports are supported.

DCC turns “cleaned data” into an explainable, traceable, rerunnable data product, giving project leads, analysts, and reviewers the same evidence chain.

Large-scale assessments may involve hundreds of thousands or millions of respondents, many items, and multiple dimensions. Traditional IRT software can become impractical because of memory use, runtime, or limited extensibility.

• Supports Rasch/1PL, PCM, RSM, 2PL, and GPCM models; unidimensional and between-item multidimensional structures; latent regression, multiple groups, case weights, and EAP scoring.
• Parallel E-steps, person-block streaming, dimension decomposition, analytical gradients, hybrid Newton updates, and automatic engine routing are designed for scale.
• Optional quadrature pruning reports measured approximation error; exact computation remains the default and never changes silently.

Grid, streaming, and automatic engines make routine analyses simple while keeping large-data execution controllable. Chinese and English manuals support handover to applied teams.

Survey-weighting scripts often fail silently when sample and target categories differ, cells are too thin, group totals drift after trimming, or multiple target sources use inconsistent schemas. A final weight column alone cannot show what was checked or which decisions were made.

• One canonical target accepts population data, weighted reference samples, or manual margins and compares sample and target before computing weights.
• A declared collapse ladder exposes incompatibilities and thin cells first, then applies reviewed category merges consistently to sample and target.
• Raking and post-stratification share one dispatcher and result contract; the guided path never bypasses a blocking precheck.
• Diagnostics, a decision ledger, bilingual reports, and audit exports make each calibration run reviewable.

WFC turns one-off weighting scripts into a standard precheck → execute → diagnose workflow, reducing the risk of weights that converge numerically while drifting from the intended population contract.