admiralagent

Spec-driven ADaM code generation with LLM-translated layers.

Idea: the LLM never writes R code. It only translates each spec variable's free-text derivation into a layer IR drawn from a closed vocabulary; a deterministic compiler then emits executable {admiral} code with # CHECK: validation comments, plus a provenance sidecar.

# Install from r-universe (prebuilt binaries for Windows/macOS/Linux):
# install.packages("admiralagent", repos = c(
#   "https://yanmingyu92.r-universe.dev", "https://cloud.r-project.org"))
# Dev version from GitHub: pak::pak("yanmingyu92/admiralagent")
spec <- mock_spec_adsl()

ir <- classify_variables(spec, "ADSL", backend = "rules")   # zero-dependency baseline
# ir <- classify_variables_llm(spec, "ADSL", chat)           # ellmer upgrade path

validate_ir(ir)                       # schema + semantic gate
cat(render_program(ir))               # executable admiral script + CHECK comments
write_artifact(ir, dir = "gen")       # script + JSON sidecar (model/IR/hash)

Generated block example:

# ---- TRTSDTM | step 2/2: impute_dtc ----------------------------------
# Spec origin: "Earliest EXSTDC among EX records with EXDOSE > 0, impute to first"
# CHECK: TRTSDTF flag exists; NA only when year component missing
adsl <- adsl |>
  admiral::derive_vars_dtm(
    new_vars_dtm = exprs(TRTSDTM),
    dtc = EXSTDTC,
    highest_imputation = "M",
    date_imputation = "first"
  )

Executing an IR is self-auditing: execute_ir() writes a hash-chained audit log with one record per variable plus one run_manifest per run, automatically — no manual manifest step. The artifact that is validated, approved and archived is the IR itself; the LLM never writes R code.

Design & architecture: see DESIGN.md. Development conventions: AGENTS.md.

See it run on real CDISC pilot data: the CDISC Pilot 5 automation showcase feeds the real pilot 5 submission spec and SDTM through the full pipeline (spec → IR → gate → render → execute → oracle comparison) with both the rules and DeepSeek LLM backends — real accuracy numbers, an explicit needs-human inventory, and an honest limitations section. Reproduce locally with Rscript demo/automation/run_all.R (report: demo/automation/REPORT.md). Two companion pieces go deeper: a controlled experiment pits the constrained-IR route against free-form code generation (same model, same spec, same oracle), and Anatomy of a Derivation follows three real variables through the full spec → IR → gate → code → oracle chain.