NEWS

---

PRA 0.7.0

• Completed the S3 method surface: mcs, smm, dsm, prob_net and pra_sigmoidal_fit objects now all have print(), summary() and plot() methods. prob_net previously had none and fell through to print.default().
• plot.prob_net() draws the network as a layered DAG in base graphics, with no optional dependency required.
• plot.pra_sigmoidal_fit() recovers the data, column names and model type from the fitted object, so plot(fit) works with no further arguments. plot_sigmoidal() is unchanged and still supported.
• validate_cor_mat() now rejects a correlation matrix that is not positive semi-definite. Such a matrix previously made smm() return a negative variance and a NaN standard deviation, and made mcs() fail inside chol() with an opaque message. mcs() additionally requires strict positive definiteness, since it factorizes the matrix.
• print.mcs() and print.smm() now return their input invisibly.
• Fixed scientific notation in MCP tool output, where values such as 200000 rendered as 2e+05 and the thousands separator was silently dropped.
• Added graphics, grDevices and tools to Imports; all three were already in use but undeclared. Dropped the unused corrplot suggestion.
• Documentation: pra_mcp_server() and the README no longer claim that every analytical function is exposed over MCP. The probabilistic network functions are not among the twelve tools the server advertises.
• Added a long-form vignette, vignette("pra", package = "PRA"), covering every analytical module with worked examples and an end-to-end case study.
• inst/CITATION now reads the version from DESCRIPTION instead of naming a hard-coded one, which had gone stale at 0.4.0.

PRA 0.6.0 (2026-08-28)

• cor_matrix() now draws column i from distribution i (previously random with replacement, so columns could repeat or omit inputs); columns are now labelled with distribution names. Results will differ.
• Correlation-matrix checks (symmetry, [-1, 1] range, unit diagonal) are now shared across mcs(), smm(), and sensitivity().
• sensitivity() warns when an index is negative, which can happen under negative correlation.
• cost_pdf() no longer requires sum(risk_probs) <= 1.
• cost_post_pdf() gained risk_probs, so unobserved (NA) risks are drawn from their prior instead of treated as absent; warns if omitted with NAs present.
• predict_sigmoidal() now errors on a model_type/fit mismatch instead of returning NA confidence bounds.

PRA 0.5.0

• mcs() fixes the Cholesky decomposition for correlated tasks — target correlations and marginal means were previously distorted. Correlated results will differ; independent tasks are unaffected.
• AI integration is now MCP-only. Removed pra_chat(), pra_app(), slash commands, and the RAG functions; use pra_mcp_server() and the bundled Agent Skills (inst/skills/) instead.
• Dropped ragnar, shiny, shinychat, bslib, vitals, mockr; added corrplot, igraph, networkD3, devtools, remotes, mcptools.
• New (experimental) probabilistic network module: prob_net(), prob_net_sim(), prob_net_learn(), prob_net_update().
• Probabilistic networks (breaking): links are now load-bearing — edges must match the distributions' dependencies, nodes must be topologically ordered, and adjacency_matrix is now directed. Networks that were already self-consistent are unaffected.
• Bayesian risk (breaking): risk_prob()/risk_post_prob() combine multiple causes with a noisy-OR, keeping results in [0, 1]. Single-cause results are unchanged.
• Sensitivity analysis (breaking): sensitivity() indices now reflect each task's actual variance contribution instead of always returning 1.0 for independent tasks.
• Added the building_project example dataset.
• Vignettes migrated to a companion Quarto book and a standalone manuscript.
• License changed from CC BY 4.0 to MIT + file LICENSE.

PRA 0.4.0 (2026-04-08)

• DSM Improvements
• Bayesian Improvements
• Agentic AI Framework
  • Slash Commands
  • Natural Language Interface
  • RAG-enhanced Reasoning
  • Local Execution via Ollama
  • Shiny App
  • Evaluation Framework

PRA 0.3.0 (2024-08-20)

• Previous changes and improvements...

PRA 0.2.0 (2024-07-02)

• Previous changes and improvements...

PRA 0.1.0

• Previous changes and improvements...