Changelog — PACER

v3.0.0

September 2026

Nexus Desktop

PACER becomes callable by AI. Every module now runs from Claude Desktop in plain language over the Model Context Protocol (MCP) — the whole psychometric engine, driven by conversation.

What's new

🔌

Connect to Claude Desktop

A single click in Settings → Connect to Claude Desktop wires PACER into Claude as an MCP server. Nothing else to install — no Node, no configuration files, no separate downloads. The connector and the analytic backend ship inside the app; PACER spawns its own engine on demand and runs every analysis locally. Your raw data stays on your machine — Claude receives only the results you ask it to interpret, never the full dataset. The Security page documents the exact data flow.

🧰

The complete engine, as 14 tools

Every module is exposed: IRT calibration, test characteristic & information curves, item & ICC plots, examinee scoring, classical item analysis, DIF, equating, automated test assembly (ATA), cognitive diagnostic models (DINA/DINO calibration + skill profiling), latent regression, explanatory IRT (LLTM), and plausible values — plus data-format helpers. Ask in natural language ("calibrate LSAT.csv with the 3PL," "assemble two parallel 40-item forms," "profile these examinees"); every tool honors PACER's exact UI defaults, so a request through Claude matches a click in the app.

🧠

PACER computes, Claude interprets

The numbers stay exact and verifiable — PACER runs the estimation; Claude narrates the results, drafts the read, and returns plot-ready data for every chart the UI shows. Rich, structured outputs (standard errors, fit statistics, convergence, warnings) keep the interpretation anchored to real diagnostics. The psychometrician stays in the loop: Claude drafts, you confirm.

Windows desktop

macOS desktop

v2.0.0

June 2026

Discover Web & Desktop

IRT Regression arrives — Latent Regression and Explanatory IRT (LLTM) bring person-level and item-level structural modeling to PACER for the first time.

What's new

📊

Latent Regression

Model the latent ability distribution as a function of person-level covariates: θ_i ~ N(X_i′γ, σ²). Supply a pre-calibrated item parameter file and a wide-format response matrix with person covariate columns. PACER estimates regression weights γ and residual population SD σ using L-BFGS with numerical standard errors, Wald z-tests, and p-values for each covariate. Categorical predictors are dummy-coded automatically. Supports binary and polytomous items (1PL / 2PL / 3PL / GRM / GPCM).

🎲

Plausible Values

After a successful Latent Regression run, generate K posterior draws of θ for each person from the individual posterior P(θ_i | data, X_i′γ, σ²). Unlike EAP point estimates, plausible values preserve measurement uncertainty for downstream group-level analyses, matching the methodology used in large-scale assessments (NAEP, PISA). Results download as CSV with one column per draw.

📈

Explanatory IRT (LLTM)

The Linear Logistic Test Model decomposes item difficulty as a linear function of observable item attributes: b_j = W_j′β. Load long-format data with item attribute columns alongside responses, select predictors (categorical or numeric), and PACER runs the EM algorithm to jointly estimate attribute weights β and population SD σ. Polytomous items (e.g. ordered rating scales) are automatically recoded using adjacent-category coding. Results include predictor importance charts, parameter estimates with SEs and p-values, predicted item difficulties, and AIC/BIC model fit statistics. Guided tours in all 8 supported languages.

Web application

Windows desktop

v1.3.1

May 2026

Anchor Upload Web & Desktop

Bulk anchor parameter upload via CSV or Excel drop zone — set dozens of fixed item parameters in one step instead of one row at a time.

What's new

⚓

Batch Anchor Parameter Upload

The IRT Calibration anchor accordion now includes a drag-and-drop file zone that accepts .csv, .xlsx, and .xls files. The file uses long format — one row per parameter — with three columns: item, param, and value. PACER validates every row against the current item selection and model configuration: unknown item names, non-numeric values, parameters invalid for the item's model (e.g. anchoring a on a 1PL item), and RSM items (which cannot be individually anchored) are all caught and reported in the validation modal. Valid rows load even when some rows fail, so partial files are usable. Duplicate item + param combinations are accepted with a warning — the last value in the file wins. A ⬇ Download template button generates a pre-filled starter CSV with all currently selected item names, ready to annotate and upload. The guided tour Step 8 (Item Anchoring) and the Template Library have both been updated with full format documentation in all eight supported languages.

Web application

Windows desktop

v1.3.0

May 2026

Multilingual Web & Desktop

Three additional languages added to the guided tour system — Korean, German, and Turkish — bringing total language coverage to eight.

What's new

🌐

Korean, German & Turkish Tour Translations

All eight guided tours are now available in Korean (한국어), German (Deutsch), and Turkish (Türkçe), joining the five languages shipped in v1.2.0. Every tour card label, tagline, description, step title, step body, and navigation button is fully translated in each new language. As with existing languages, the selection persists across sessions via localStorage and requires no configuration changes. The additions are purely strings — the tour engine itself is unchanged. Total coverage: English, Spanish, Portuguese, French, Simplified Chinese, Korean, German, Turkish.

Web application

Windows desktop

v1.2.0

May 2026

Global Web & Desktop

Full internationalization of the guided tour system — five languages, zero configuration, language preference persists across sessions.

What's new

🌐

Guided Tour Internationalization

All eight guided tours — Item Analysis, Item Calibration, Test Scoring, CDM Calibration, Skill Profiler, Equating, Test Construction, and DIF Analysis — are now available in five languages: English, Spanish, Portuguese (Brazilian), French, and Simplified Chinese. Every tour card label, tagline, description, step title, step body, and navigation button is fully translated. Language is selected from a dropdown in the tour menu and remembered across page reloads via localStorage. The architecture uses a zero-dependency strings file (pacer-tour-strings.js) that is loaded independently of the tour engine, making future language additions a strings-only change.

Web application

Windows desktop

Internationalization Spanish Portuguese French Simplified Chinese Guided Tours

v1.1.0

April 2026

Sidecar Desktop Only

Brings a local AI assistant on board — Gemma 4B runs directly inside the PACER Desktop app, no internet required.

What's new

🤖

Local AI Assistant

Gemma 4 Billion runs entirely on-device within the PACER Desktop app, powering an AI assistant for guidance and interpretation of psychometric results. No data leaves your machine. This feature is exclusive to the Windows desktop client — the web application is unaffected.

Windows desktop

AI Assistant Gemma 4B On-Device Desktop Only

v1.0.0

April 2026

Kickstart Initial Launch

The first public release of PACER — a fully featured psychometric computing platform for researchers and testing organizations, available on the web and desktop.

What's included

📊

Item Analysis

Classical test theory statistics including p-values, point-biserial correlations, distractor analysis, and reliability estimates.

⚙️

Item Calibration

Full IRT calibration engine supporting 1PL, 2PL, 3PL, GRM, GPCM, PCM, and RSM via marginal maximum likelihood with EM and Gauss-Hermite quadrature.

🎯

Scoring

IRT-based ability estimation with EAP, MAP, and MLE methods. Full support for polytomous and mixed-format tests.

🧩

Cognitive Diagnostic Models

DINA, DINO, and G-DINA calibration and scoring for diagnostic classification and skill mastery profiling.

🔗

Equating & Linking

Six equating methods including Stocking-Lord, Haebara, Mean/Sigma, and Mean/Mean for linking test forms to a common scale.

🏗️

Test Construction & Assembly

Automated test assembly powered by Google OR-Tools CP-SAT solver. Blueprint constraints, TIF optimization, item groups, and simultaneous multi-form assembly.

🔍

DIF Analysis

Differential item functioning detection using Mantel-Haenszel and standardized mean difference statistics across focal and reference groups.

👥

Multigroup Calibration

Bock-Zimowski multigroup IRT for concurrent calibration across multiple populations with group-specific ability distributions.

Web application

Windows desktop

REST API

IRT CDM CTT DIF Equating Test Assembly Multigroup PDF Reports PACER University