LLM-assisted content analysis for the social sciences, built on LLMR. Three connected concerns, one package: coding, robustness audits, and replication archives. When a text label becomes a variable in quantitative analysis, this is the workflow that keeps the measurement honest.
The three workflows
Coding turns a codebook and a sealed gold standard into error-corrected category prevalences:
g <- gold_set(data, text = "text", labels = "label") # sealed test split
p <- protocol_lock(protocol(codebook, config)) # hash-locked instrument
v <- validate_protocol(p, g) # the one honest evaluation
coded <- code_corpus(corpus, p, "text") # apply to the full corpus
gc <- gold_correct(coded, g) # corrected prevalences + SEsRobustness audits ask whether a coded conclusion survives the measurement multiverse of prompts, models, label orders, and temperatures:
plan <- audit_plan(data, "text", estimator, labels, prompt)
plan <- audit_add_models(plan, configs)
plan <- audit_add_perturbations(plan, label_order = c("as_given", "reversed"),
temperature = c(0, 0.7))
a <- audit_run(plan)
audit_stability(a); audit_fragility(a) # sign, rank, fragilityArchives turn the audit log that LLMR writes into a reviewer-runnable replication record:
ar <- archive_build(log) # content-addressed
ar <- archive_seal(ar) # frozen under a root hash
archive_check(ar) # integrity + completeness
replay <- archive_replay(ar) # recompute offline, no keysThe shared generics LLMR::diagnostics(), LLMR::report(), and tibble::as_tibble() dispatch across all three families of result objects.
Point-and-click
The same three workflows run from a Shiny GUI, for collaborators who would rather not write R:
install_gui_deps() # shiny, bslib, DT, and the LLMR.shiny substrate
run_content_studio() # coding, robustness audit, and archive tabsThe GUI wraps the package API rather than reimplementing it, reads provider keys from environment variables only, and has a deterministic demo mode that runs offline.