Skip to contents

Changelog

Source: NEWS.md

admiralagent 0.1.1

CRAN resubmission: addresses the first-review comments and folds in the post-0.1.0 development features.

  • Description now expands all acronyms (ADaM, LLM, IR, P21, CDISC) and references the CDISC Analysis Data Model standard.
  • Removed the example from the unexported parse_define(); every exported function now carries a small executable example (LLM- and MCP-server-only functions keep \dontrun{} with the reason stated: provider API key or blocking stdio server).
  • write_artifact() / write_program_artifact() no longer default dir: an explicit output directory is required, so no function writes into the user’s working directory by default; examples and tests write to tempdir().
  • Removed the remaining <<- assignments (closure accumulators rewritten over explicit local environments; no behaviour change, full test suite green).
  • New accept_model(): model+harness acceptance gate. A candidate LLM runs the spec-to-IR regression corpus and is cross-compared against the rules baseline at full canonical semantics; both thresholds are required arguments (acceptance criteria must be prespecified), any pipeline error rejects the candidate, and the decision is logged as a digest-only model_acceptance audit event.
  • New routing_plan(): derives the execution routing directly from spec lineage — topological waves, per-variable dependencies, and the two fail-closed lanes (cycle-blocked and human-routed variables stay waveless). The ordering is computed, not declared, replacing a hand-registered process DAG.
  • New classify_variables_dual(): runs the rules classifier and the LLM pipeline independently on the same spec, compares the two translations per variable at full canonical semantics (ir_fingerprint(level = "full")), adopts the rules IR on agreement, and escalates any divergence (disagree / rules_abstain / llm_abstain) to needs_human — no side is ever silently preferred. The decision is recorded as a digest-only dual_consensus audit event; dual_disagreements() extracts the human review queue.

admiralagent 0.1.0

Initial release.

  • Translate P21-style ADaM specifications into layer IR over a closed vocabulary, then deterministically compile the IR into admiral/metatools programs with # CHECK: validation comments and provenance sidecars (classify_variables(), render_program(), write_artifact()).
  • Zero-dependency rules classifier (backend = "rules") as the default baseline; optional LLM translation (backend = "llm", via ‘ellmer’) emits IR only — the LLM never writes R code and every model response is validated against the closed vocabulary before use.
  • Sandbox gate for executing generated programs, hash-chained audit log, dataset-JSON reader, define.xml (2.0) metadata reader, and dependency/impact reporting across ADaM datasets.
  • Study-level drivers: execute_one(), execute_study() with derivation ordering, gates, and manifest generation.