Review Follow-up Status

この文書は、公開リポジトリ化前後の厳しめレビュー指摘に対する現在の対応状況を記録します。 古い実装指示をそのまま残すと、未解決問題と完了済み機能が混ざって第三者レビューを混乱させるため、 ここでは done / tracked / external を明確に分けます。

このリポジトリは workflow fixture generator です。PK truth source、臨床推論、投与設計、 規制提出用モデル妥当化の証拠としては扱いません。

Current Triage

Review item Status Current evidence Remaining action
IV infusion handling Done tools/run_demo_set.py uses infusion_h for IV infusion equations; tools/make_analysis_inputs.py emits PopPK RATE; tests cover both paths Keep regression tests when changing regimen logic
t1/2 vs CL/V structural mismatch labeling Done, now machine-readable in targets.yml and workflow manifests targets.yml carries structural_mismatch.acknowledged; run-level MANIFEST.yml includes detected_structural_mismatch, acknowledged_structural_mismatch, relative_error, and attainability_status Do not auto-fix PK values; use source review before changing canonical data
AUC target circularity Done, now machine-readable in targets.yml and workflow manifests targets.yml carries basis, target_basis, independent_literature_target, source_value, and role; run-level MANIFEST.yml mirrors these under target_metadata.auc Replace targets.auc only when a literature AUC has source/raw text/unit conversion evidence
BLQ / LLOQ fixture support Done for fixture contract assay.lloq creates SDTM-like BLQ flags and PopPK BLQ/CENS/LIMIT; tests cover SDTM-like and analysis inputs External M3 likelihood execution remains tool/environment-specific
Demo IIV/residual meaning Done docs/SCHEMA.md and docs/USER_GUIDE.md distinguish demo-only variability from model-specific external runner behavior Revisit only if full stochastic model execution is added
Predose convention Done Default predose remains DV=0/MDV=0; --predose-mdv1 is available and tested Site-specific PopPK adapters may still choose another convention
Covariate / absorption scope Done Docs state demographic covariates are fixture attributes, not a justified covariate model Add model-specific covariate support only behind explicit schema/tests
SC/IM route guard Done Demo generator treats SC/IM as first-order absorption and rejects unknown routes instead of silently using bolus Do not broaden route support without tests
Interpolation / sorting intent Done linear and log-linear behavior is documented; tests cover log-linear sampling Formal NCA interpolation remains downstream-tool scope
Package vs checkout UX Tracked README states this is a git-checkout tool; wheel/sdist ship code only and not the drug library Decide separately if the project should become a fully installable data package
Value-level provenance Tracked pk.yml stores sources/raw/parsed/derived, but value-to-source snippets and reviewer status are not uniformly normalized Improve one drug at a time with source/raw text/unit conversion/reviewer evidence
External Phoenix/NONMEM/nlmixr2 execution External Probe/smoke paths exist; real execution needs licensed/local environments Track in docs/READINESS_GAPS.md and release notes
Independent README-only user test External Checklist/template exists Needs a third-party tester, not Codex

Machine-readable Fields

Run-level workflow manifests now carry target provenance and structural mismatch metadata:

target_metadata:
  parameter_pair_policy: spec_theta_uses_pk_yml_derived_cl_v_abs
  auc:
    basis: dose_over_cl
    target_basis: dose_over_cl_not_literature_auc
    independent_literature_target: false
  t_half:
    basis: literature_target_retained_as_check
    detected_structural_mismatch: true
    acknowledged_structural_mismatch: true
    attainability_status: WARN
    relative_error: 0.406

These fields are run metadata. They do not rewrite pk.yml, targets.yml, or spec_pk1_*.yml.

Rules For Future Fixes

  • Do not change PK numeric values without source text, conversion formula, units, and reviewer rationale.
  • Treat Dose/CL AUC as an integration consistency target, not as independent literature validation.
  • Treat detected_structural_mismatch: true as a computed fixture limitation, not as permission to silently recalibrate CL, V, or t1/2.
  • Treat acknowledged_structural_mismatch: true as a human-reviewed fixture policy label, not as clinical validation.
  • Keep external tool execution claims separate from parser/probe smoke checks.