pharmaskills 0.6.0__py3-none-any.whl

pharmaskills/__init__.py

@@ -0,0 +1,6 @@
+"""PharmaSkills: pharma skills for AI agents."""
+
+from pharmaskills._meta import DISPLAY_NAME, NAME, VERSION
+
+__all__ = ["DISPLAY_NAME", "NAME", "VERSION", "__version__"]
+__version__ = VERSION

pharmaskills/__main__.py

@@ -0,0 +1,4 @@
+from pharmaskills.cli import main
+
+if __name__ == "__main__":
+    main()

pharmaskills/_meta.py

@@ -0,0 +1,11 @@
+"""Single source of truth for the product identity (PRD s22).
+
+A rename is one package-directory move plus ``[project] name`` in
+``pyproject.toml`` and ``name`` in ``.claude-plugin/plugin.json``.
+"""
+
+NAME = "pharmaskills"
+DISPLAY_NAME = "PharmaSkills"
+VERSION = "0.6.0"
+TAGLINE = "Pharma skills for AI agents."
+USER_AGENT = f"{NAME}/{VERSION} (+https://github.com/aiagentpro/{NAME})"

pharmaskills/_skills/adam-spec/SKILL.md

@@ -0,0 +1,138 @@
+---
+name: adam-spec
+description: Specify ADSL and ADAE from SDTM datasets and the SAP, then build the datasets, derivation programs and spec workbook with unit-tested derivations. Use for ADaM specification, ADSL or ADAE derivation, TRTEMFL or population flags.
+license: Apache-2.0
+metadata:
+  stage: clinical-data
+  tier: 1
+  standards: [ADaMIG 1.3, ADaM OCCDS 1.0, SDTMIG 3.4, ICH E9]
+  tools:
+    - pharmaskills adam build
+    - pharmaskills adam test
+    - pharmaskills docs banner
+  inputs: [sdtm_xpt_datasets, sap, study_yaml]
+  outputs: [adam_spec_yaml, adam_spec_xlsx, adsl_xpt, adae_xpt, adsl_py, adae_py, adam_build_report_json, adam_test_report_json]
+  gxp_note: Draft for qualified review. Not a validated system and not 21 CFR Part 11 compliant.
+  min_pharmaskills: "0.3.0"
+---
+
+# ADaM specification (ADSL and ADAE)
+
+## Purpose
+
+Produce a reviewable specification and the derived ADSL and ADAE analysis datasets for a
+statistical programmer, starting from SDTM datasets and the statistical analysis plan (SAP).
+You translate SAP definitions into a declarative spec (never into data), the deterministic
+`pharmaskills adam` engine derives the datasets, and `pharmaskills adam test` proves the key
+derivations. Everything is a draft for qualified human review.
+
+## When to use / when not to use
+
+Use to specify or derive ADSL (population flags, treatment variables, treatment dates, age
+groups) and ADAE (treatment-emergent flag, analysis dates, severity, duration, first
+occurrence), or to check existing ADSL and ADAE derivations against the SAP.
+
+Do not use for: any other ADaM dataset (ADTTE, ADLB, ADVS, ADEFF and all BDS parameter
+datasets are out of scope in v1; say so and offer a specification outline marked unverified),
+SDTM mapping (use `sdtm-mapper`), Define-XML (use `define-xml`), or real patient data that has
+not been de-identified. Never hand-edit a derived dataset or a flag value.
+
+## Inputs
+
+Required: SDTM XPT datasets DM, EX and AE (a folder of `.xpt` files or a study folder with
+`out/sdtm/xpt`, `gold/sdtm` or `sdtm`), and the SAP (DOCX or Markdown) sections on analysis
+populations, treatment definitions, treatment-emergent adverse events and subgroups.
+
+Ask once for anything missing, then proceed with stated assumptions and `[DATA NEEDED: ...]`
+placeholders. Typical gaps: follow-up window length, partial date rule, definition of the
+safety population, age band cut points.
+
+State the region assumed (US FDA with ICH unless told otherwise). If `pharmaskills` is not
+installed, tell the user to run `uv tool install pharmaskills`, check `pharmaskills --version`
+is at least `min_pharmaskills`, and produce only the spec outline and notes, clearly marked
+unverified, with no datasets.
+
+## Workflow
+
+You are the model in steps 1, 2 and 6. Everything else is a deterministic CLI call.
+
+1. **Read the SAP.** Extract the definitions listed in `references/derivations.md` (population,
+   TRT01P and TRT01A, TRTSDT and TRTEDT, treatment-emergent window, partial and missing date
+   rules, severity coding, age groups). Quote the SAP section for each. Where the SAP is
+   silent, choose the documented default, record it as an assumption and flag it.
+2. **Write `adam_spec.yaml`** from `templates/adam_spec.yaml`: arms keyed by ARMCD, age bands,
+   `followup_days`, `partial_start` (`conservative`, `earliest` or `latest`),
+   `missing_start_is_emergent`, `safety_min_dose`. Never put data values in the spec.
+3. **Build.** `pharmaskills adam build <sdtm-or-study> --spec adam_spec.yaml --out out/`
+   writes `adsl.xpt`, `adae.xpt`, Dataset-JSON, `adam_spec.xlsx`, `adsl.py`, `adae.py` and
+   `adam_build_report.json`. The same inputs give byte-identical files.
+4. **Test.** `pharmaskills adam test <sdtm-or-study> --spec adam_spec.yaml --expected
+   expected_values.yaml --out out/adam_test_report.json`. Built-in checks cover one row per
+   subject, SAFFL from exposure, TRTEMFL against an independent recomputation and record counts.
+   Build `expected_values.yaml` from the raw SDTM by hand for known cases (pre-treatment event,
+   partial start date, missing start date, event inside the follow-up window); see
+   `templates/expected_values.yaml`. Never copy values from the build output into it.
+5. **Repair loop (maximum 3 iterations).** On a failure find the cause in the spec or the SAP
+   reading, change the spec, rebuild and retest. Never change a dataset value or weaken a check.
+6. **Document.** Write `out/adam_spec_notes.md` from `templates/adam-spec-notes.md`: decisions
+   with SAP citations, tool counts, assumptions, open items. Insert the draft banner and a
+   provenance block with `pharmaskills docs banner --apply out/adam_spec_notes.md`.
+7. **Report** using the output contract below.
+
+## Output contract
+
+Under `out/`:
+
+| File | Content |
+| --- | --- |
+| `adam_spec.yaml` | Declarative spec filled from the SAP |
+| `adam_spec.xlsx` | Variable list (dataset, label, type, source, derivation) and parameters |
+| `adsl.xpt`, `adae.xpt`, `adsl.json`, `adae.json` | XPT v5 and Dataset-JSON v1.1 datasets |
+| `adsl.py`, `adae.py` | Readable derivation programs that call the tested library |
+| `adam_build_report.json` | Counts, flags, input and output hashes |
+| `adam_test_report.json` | One entry per derivation check with pass or fail |
+| `adam_spec_notes.md` | Decisions, assumptions, open items, banner and provenance |
+
+Dates are numeric SAS dates in XPT and ISO 8601 strings in Dataset-JSON. State plainly in the
+answer: the scope (ADSL and ADAE only), record counts, flag counts, check results and
+every assumption.
+
+## Verification
+
+Run and report pass or fail for each: `adam build` (exit 0), `adam test` (exit 0, all checks),
+`pharmaskills docs banner` applied to the notes. Acceptance on the synthetic benchmark PSK-001:
+ADSL has one row per subject (138), SAFFL is Y for the 120 treated subjects, TRTEMFL matches
+the independent recomputation for all 168 AE records, rebuild is byte-identical. On failure
+fix the spec and rerun up to three times, then report what remains. No CDISC CORE ADaM run is
+performed by this skill; say so and do not describe the datasets as conformant.
+
+## Guardrails
+
+- Insert the draft banner in every generated document; never claim the datasets are validated,
+  submission-ready or CORE-clean.
+- Never fabricate data, subject values, dates, flags or SAP content. Missing SAP rules become
+  `[DATA NEEDED: ...]` placeholders plus a flagged assumption.
+- Never write or edit ADaM records directly and never run model-generated code over the data.
+  Only the spec changes; the engine executes it.
+- Refuse requests to set flags to improve a safety picture, to hide events or to bypass the
+  SAP definition; offer a documented sensitivity analysis instead.
+- ADSL and ADAE only; state that other ADaM datasets are out of scope in v1.
+- Warn when inputs look like real patient data; `pharmaskills docs detect-pii` checks this.
+- Every claim about a guideline cites identifier and section (see `references/`).
+
+## References
+
+- `references/derivations.md` - supported derivations, rules, defaults and limits of v1
+- `references/adam-notes.md` - ADaMIG 1.3 and OCCDS 1.0 notes with section citations
+- `references/spec-format.md` - the YAML spec and expected-values file formats
+- Public sources: CDISC ADaM Implementation Guide v1.3 and ADaM OCCDS v1.0 (cdisc.org/standards),
+  ICH E9 (database.ich.org), CDISC SDTMIG 3.4.
+
+Templates: `templates/adam_spec.yaml`, `templates/expected_values.yaml`,
+`templates/adam-spec-notes.md`. Script: `scripts/run_adam.sh` (thin wrapper around
+`pharmaskills adam build` and `adam test`).
+
+## Examples
+
+Worked example with inputs and outputs: `examples/clinical-data/adam-spec/`. Synthetic inputs:
+`skills/adam-spec/evals/fixtures/` (DM, EX and AE of PSK-001, a SAP excerpt, expected values).

pharmaskills/_skills/adam-spec/references/adam-notes.md

@@ -0,0 +1,20 @@
+# ADaM notes with citations
+
+Written for this project; summarises public CDISC and ICH guidance, no standard text is copied.
+
+- ADaMIG v1.3, section 3 (ADSL): one record per subject, population indicator flags such as
+  SAFFL and ITTFL carry Y or N, treatment variables TRTxxP and TRTxxA with numeric
+  counterparts, treatment dates TRTSDT and TRTEDT. Verify section numbers against the IG you
+  hold. https://www.cdisc.org/standards/foundational/adam
+- ADaM OCCDS v1.0, section 3 (occurrence data structure): ADAE is an OCCDS dataset with one
+  row per reported occurrence, analysis dates ASTDT and AENDT, relative days ASTDY and AENDY,
+  and TRTEMFL as the treatment-emergent analysis flag. https://www.cdisc.org/standards/foundational/adam
+- ADaMIG v1.3, section 4 (date and time variables): analysis dates are numeric dates with a
+  date format; imputation flags (ASTDTF, AENDTF) record the imputed component. The engine
+  writes numeric SAS dates in XPT and ISO 8601 in Dataset-JSON.
+- ICH E9 section 5.2 (analysis sets): the analysis populations come from the protocol and SAP,
+  so population flags must be traceable to a SAP definition. https://database.ich.org/sites/default/files/E9_Guideline.pdf
+- ICH E3 section 12.2 (adverse events): displays of adverse events use the treatment-emergent
+  definition stated in the SAP. https://database.ich.org/sites/default/files/E3_Guideline.pdf
+- CDISC CORE (open-source conformance engine) is the intended place to run ADaM conformance
+  rules; the native checks in `adam test` are derivation unit tests, not conformance rules.

pharmaskills/_skills/adam-spec/references/derivations.md

@@ -0,0 +1,39 @@
+# Supported derivations in `pharmaskills adam` v1
+
+Written for this project. It summarises what the engine does, not the CDISC standard text.
+Scope: ADSL and ADAE only. ADTTE, ADLB, ADVS, ADEFF and every other BDS dataset are out of
+scope; there is no baseline, visit windowing, imputation of treatment dates, multi-period
+treatment, worsening-based emergence or time-to-event derivation in v1.
+
+## ADSL (one row per DM subject)
+
+| Variable | Rule | Spec keys |
+| --- | --- | --- |
+| TRT01P, TRT01PN | Lookup of ARMCD in `arms`; null when ARMCD is in `not_randomised_armcds` | `arms`, `not_randomised_armcds` |
+| TRT01A, TRT01AN | Lookup of ACTARMCD in `arms` for subjects with exposure | `arms` |
+| TRTSDT, TRTEDT | Min EXSTDTC and max EXENDTC over EX records with complete dates and EXDOSE at or above the threshold | `safety_min_dose` |
+| TRTDURD | TRTEDT minus TRTSDT plus 1 | none |
+| SAFFL | Y when the subject has at least one qualifying EX record, else N | `safety_min_dose` |
+| ITTFL | Y when ARMCD is a randomised arm of the spec | `arms`, `not_randomised_armcds` |
+| AGEGR1, AGEGR1N | First inclusive band containing AGE | `age_bands` |
+
+Placebo exposure has EXDOSE 0, so the default threshold 0 (greater than or equal) counts it.
+EX records with a partial start date or a missing dose do not count as exposure; report them.
+
+## ADAE (one row per AE record, subject variables merged from ADSL)
+
+| Variable | Rule |
+| --- | --- |
+| ASTDT, ASTDTF | AESTDTC as a date. Partial dates: `conservative` imputes the first dose date when it lies inside the partial period, else the first day; `earliest` the first day; `latest` the last day. Flag D (day) or M (month and day) |
+| AENDT, AENDTF | AEENDTC as a date; partial dates per `partial_end`; null when ongoing |
+| ASTDY, AENDY | Date minus TRTSDT, plus 1 when on or after TRTSDT (no day 0) |
+| ADURN, ADURU | AENDT minus ASTDT plus 1 days |
+| AESEVN | `ae.severity` map (MILD 1, MODERATE 2, SEVERE 3) |
+| TRTEMFL | Y when TRTSDT <= ASTDT <= TRTEDT + `followup_days`; missing ASTDT follows `missing_start_is_emergent`; null when the subject was not treated |
+| AOCCFL | Y on the first treatment-emergent record per subject by ASTDT then AESEQ |
+
+## Assumptions to state every time
+
+Follow-up window length, partial date rule, missing start rule, severity coding, safety
+population definition and the region (US FDA with ICH by default). Each must be traceable to a
+SAP section or flagged as an assumption.

pharmaskills/_skills/adam-spec/references/spec-format.md

@@ -0,0 +1,29 @@
+# Spec and expected-values formats
+
+## `adam_spec.yaml` (validated by `pharmaskills adam build`; unknown keys are errors)
+
+```yaml
+study: PSK-001
+arms: {PBO: {trt: Placebo, n: 1}, ACT: {trt: Active 10 mg, n: 2}}
+not_randomised_armcds: [SCRNFAIL, NOTASSGN, NOTTRT]
+safety_min_dose: 0.0
+age_bands: [{label: "<45", n: 1, high: 44}, {label: ">=45", n: 2, low: 45}]
+ae: {followup_days: 30, partial_start: conservative, partial_end: latest,
+     missing_start_is_emergent: true, severity: {MILD: 1, MODERATE: 2, SEVERE: 3}}
+```
+
+## `expected_values.yaml` (read by `pharmaskills adam test --expected`)
+
+```yaml
+adsl:
+  subjects:                       # per subject, any ADSL variable; dates as ISO strings
+    PSK-001-101-001: {SAFFL: Y, TRT01A: Active 10 mg, TRTSDT: "2025-02-15"}
+  counts: {SAFFL: {Y: 120, N: 18}}
+adae:
+  records:                        # per USUBJID and AESEQ, any ADAE variable
+    - {USUBJID: PSK-001-101-007, AESEQ: 1, TRTEMFL: null}
+  counts: {TRTEMFL: {Y: 161}}
+```
+
+Exit codes: 0 all checks passed, 1 input or spec error, 2 at least one check failed (the report
+JSON lists every check with its detail).

pharmaskills/_skills/adam-spec/scripts/run_adam.sh

@@ -0,0 +1,14 @@
+#!/bin/sh
+# Thin wrapper: build ADSL and ADAE, then run the derivation tests.
+# Usage: run_adam.sh <sdtm-or-study-folder> <adam_spec.yaml> [expected_values.yaml] [out-dir]
+set -eu
+sdtm="${1:?usage: run_adam.sh <sdtm-or-study-folder> <adam_spec.yaml> [expected.yaml] [out]}"
+spec="${2:?missing adam_spec.yaml}"
+expected="${3:-}"
+out="${4:-out}"
+pharmaskills adam build "$sdtm" --spec "$spec" --out "$out"
+if [ -n "$expected" ]; then
+  exec pharmaskills adam test "$sdtm" --spec "$spec" --expected "$expected" \
+    --out "$out/adam_test_report.json"
+fi
+exec pharmaskills adam test "$sdtm" --spec "$spec" --out "$out/adam_test_report.json"

pharmaskills/_skills/adam-spec/templates/adam-spec-notes.md

@@ -0,0 +1,25 @@
+# ADSL and ADAE derivation notes for <STUDYID>
+
+Region assumed: <US FDA with ICH | other>. Source: <SDTM datasets and SAP version>.
+Scope: ADSL and ADAE only (v1).
+
+## Results of `pharmaskills adam build`
+
+| Item | Value |
+| --- | --- |
+| ADSL records | <from adam_build_report.json> |
+| ADAE records | <from adam_build_report.json> |
+| SAFFL = Y | <from adam_build_report.json> |
+| TRTEMFL = Y | <from adam_build_report.json> |
+
+## Derivation decisions
+
+- <variable>: <rule> (SAP section <n>; ADaMIG 1.3 section <n>)
+
+## Verification (`pharmaskills adam test`)
+
+<n of n> checks passed: <list>.
+
+## Open items
+
+- [DATA NEEDED: <SAP rule or input that was missing>]

pharmaskills/_skills/adam-spec/templates/adam_spec.yaml

@@ -0,0 +1,18 @@
+# ADaM v1 spec for `pharmaskills adam build` and `adam test` (ADSL and ADAE only).
+# Fill every value from the SAP; leave a [DATA NEEDED: ...] comment for anything the SAP omits.
+study: PSK-001                    # STUDYID from the SDTM data
+arms:                             # keyed by SDTM ARMCD / ACTARMCD; SAP section on treatments
+  PBO: {trt: Placebo, n: 1}
+  ACT: {trt: Active 10 mg, n: 2}
+not_randomised_armcds: [SCRNFAIL, NOTASSGN, NOTTRT]
+safety_min_dose: 0.0              # EXDOSE >= this value counts as exposure (placebo is 0)
+age_bands:                        # SAP section on subgroups; null low or high = open end
+  - {label: "<45", n: 1, high: 44}
+  - {label: "45-64", n: 2, low: 45, high: 64}
+  - {label: ">=65", n: 3, low: 65}
+ae:
+  followup_days: 30               # treatment-emergent window after TRTEDT (SAP definition)
+  partial_start: conservative     # conservative | earliest | latest
+  partial_end: latest             # earliest | latest
+  missing_start_is_emergent: true
+  severity: {MILD: 1, MODERATE: 2, SEVERE: 3}

pharmaskills/_skills/adam-spec/templates/expected_values.yaml

@@ -0,0 +1,13 @@
+# Hand-built expected values for `pharmaskills adam test --expected`.
+# Compute every value from the raw SDTM records yourself. Never copy them from a build output.
+adsl:
+  subjects:
+    # USUBJID: {SAFFL: Y, TRT01A: "<arm text>", TRTSDT: "YYYY-MM-DD", TRTEDT: "YYYY-MM-DD"}
+  counts:
+    # SAFFL: {Y: <n>, N: <n>}
+adae:
+  records:
+    # - {USUBJID: "<id>", AESEQ: <n>, TRTEMFL: Y}      # on treatment
+    # - {USUBJID: "<id>", AESEQ: <n>, TRTEMFL: null}   # starts before first dose
+  counts:
+    # TRTEMFL: {Y: <n>}

pharmaskills/_skills/aggregate-report/SKILL.md

@@ -0,0 +1,115 @@
+---
+name: aggregate-report
+description: Draft a PBRER (ICH E2C(R2)) or DSUR (ICH E2F) section structure with data placeholders, a stated reporting interval and a structured benefit-risk framework; checks mandatory sections. Never invents exposure or conclusions. Draft for qualified review.
+license: Apache-2.0
+metadata:
+  stage: safety
+  tier: 2
+  standards: [ICH E2C(R2), ICH E2F, EU GVP Module VII]
+  tools:
+    - pharmaskills docs check-structure --standard pbrer
+    - pharmaskills docs check-structure --standard dsur
+    - pharmaskills docs build
+    - pharmaskills docs banner
+  inputs: [report_type, reporting_interval, data_summaries, clinical_tables, nonclinical_summaries]
+  outputs: [pbrer_docx, dsur_docx, structure_report_json]
+  gxp_note: Draft for qualified review. Not a validated system and not 21 CFR Part 11 compliant.
+  min_pharmaskills: "0.4.0"
+---
+
+# Aggregate safety report drafter
+
+## Purpose
+
+Produce the section skeleton of a periodic benefit-risk evaluation report (PBRER) or a
+development safety update report (DSUR) in the order of the ICH format, with the reporting
+interval stated, supplied data filled in with their sources, everything else a placeholder, and
+a structured benefit-risk evaluation framework. It is a drafting aid for the safety physician
+and the qualified person responsible for pharmacovigilance. It does not assess causality,
+conclude on benefit-risk or replace their review and the sponsor's reporting obligations and
+timelines.
+
+## When to use / when not to use
+
+Use at the start of an aggregate report cycle, to get a compliant structure with the available
+data placed and the gaps listed.
+
+Do not use to write the benefit-risk conclusion, to select signals, to compute exposure
+without supplied dosing or sales data, or to file the report. Do not choose the report type for
+the user: a DSUR covers an investigational product in development, a PBRER an authorised
+product (and the periodicity comes from the authority list or agreement).
+
+## Inputs
+
+Required: report type (PBRER or DSUR), reporting interval (start, end, data lock point) and the
+international birth date (IBD for PBRER, DIBD for DSUR). Optional: clinical tables with
+denominators, nonclinical summaries, signal tables, exposure estimates, risk management plan.
+
+Ask once for anything missing, then proceed with stated assumptions and
+`[DATA NEEDED: ...]` placeholders. Never invent dates, exposure, counts or conclusions.
+
+State the region assumed (default: US FDA and EU in ICH format; say what regional appendices
+are not covered).
+
+If the `pharmaskills` package is not installed, say `uv tool install pharmaskills`, draft from
+the template and mark the structure check as not run.
+
+## Workflow
+
+1. **Confirm the report type and interval.** Write `Reporting interval: <start> to <end>`
+   (PBRER) or `Reporting period: <start> to <end>` (DSUR) in the first paragraph.
+2. **Create all sections in order** from `templates/pbrer-outline.md` or
+   `templates/dsur-outline.md`. Do not renumber or merge; mark not-applicable sections
+   "Not applicable" with a reason, never delete them.
+3. **Place supplied data** with `[Source: <file>]` and exact denominators. Everything else is
+   `[DATA NEEDED: <what and from whom>]`.
+4. **Benefit-risk section**: fill the framework table (`references/benefit-risk-framework.md`)
+   with evidence, source and uncertainty; assessment stays a placeholder for the safety
+   physician.
+5. **Build**: `pharmaskills docs build src.md --out pbrer.docx --skill aggregate-report`
+   (inserts the draft banner and provenance block).
+6. **Check**: `pharmaskills docs check-structure pbrer.docx --standard pbrer --out
+   pbrer_structure_report.json` (or `--standard dsur`).
+
+## Output contract
+
+| File | Content |
+| --- | --- |
+| `pbrer.docx` or `dsur.docx` (plus `.md`) | All numbered sections, interval stated, benefit-risk table, banner first, provenance last |
+| `<type>_structure_report.json` | Section presence report from the tool |
+
+Every number traces to a supplied file; every placeholder names the missing item. Not-applicable
+sections state why.
+
+## Verification
+
+Run and report: `docs check-structure` exit code (0 expected), counts present and missing
+required, that the interval line is present and matches the supplied dates, and the banner.
+Recompute one count from a supplied table by hand and state it matches. If the tool is not
+available, compare the headings to `references/pbrer-sections.md` by eye and say it was manual.
+
+## Guardrails
+
+- Insert the draft banner in every generated document.
+- Never fabricate exposure, case counts, signals, citations, PMIDs or NCT ids; use
+  `[DATA NEEDED: ...]`.
+- Never write the benefit-risk conclusion or a statement that the balance is favourable; that
+  belongs to the qualified person after medical review.
+- Do not infer causality from imbalance in counts; describe counts only.
+- State the region and the report type assumed; the format differs between PBRER and DSUR.
+- The report does not replace the sponsor's reporting obligations and timelines.
+
+## References
+
+- `references/pbrer-sections.md` - PBRER (E2C(R2)) section order and what each needs
+- `references/dsur-sections.md` - DSUR (E2F) section order and what each needs
+- `references/benefit-risk-framework.md` - the structured evaluation table
+- Public sources: ICH E2C(R2) and ICH E2F (ich.org), EU GVP Module VII (EMA).
+
+Templates: `templates/pbrer-outline.md`, `templates/dsur-outline.md`. Script:
+`scripts/check_aggregate.sh`.
+
+## Examples
+
+Worked example, both report types, from synthetic data:
+`examples/safety/aggregate-report/`. Fixtures: `skills/aggregate-report/evals/fixtures/`.

pharmaskills/_skills/aggregate-report/references/benefit-risk-framework.md

@@ -0,0 +1,16 @@
+# Benefit-risk evaluation framework
+
+A structure, not a conclusion. Rows are filled only from supplied evidence; the assessment cell
+is written by the qualified person after medical review.
+
+| Dimension | What goes in Evidence | Typical sources |
+| --- | --- | --- |
+| Condition and unmet need | Disease burden, current options | Indication dossier, literature |
+| Key benefits | Efficacy results with effect size and interval | CSRs, TLFs |
+| Key risks | Identified and potential risks with frequency and severity | Safety tables, signals |
+| Risk management | Minimisation measures and their effectiveness | RMP, label |
+| Overall conclusion | Left to the qualified person | Safety physician |
+
+For each row record the source file and the main uncertainty (small database, short duration,
+missing comparator, no causality assessment). Do not weigh benefits against risks in the draft;
+state the data and the open questions.

pharmaskills/_skills/aggregate-report/references/dsur-sections.md

@@ -0,0 +1,16 @@
+# DSUR section order (ICH E2F)
+
+Notes written for this project from the public table of contents; no guideline text copied.
+Machine-readable list: `src/pharmaskills/standards/data/dsur.yaml` (UNSURE: subsection
+numbering under sections 6 to 8 and 18; verify against the published PDF).
+
+1 Introduction. 2 Worldwide marketing approval status. 3 Actions taken for safety reasons. 4
+Changes to reference safety information. 5 Inventory of trials. 6 Estimated cumulative exposure.
+7 Line listings and summary tabulations. 8 Significant findings from clinical trials (8.1 to
+8.5). 9 Non-interventional studies. 10 Other clinical trial or study safety information. 11
+Marketing experience. 12 Non-clinical data. 13 Literature. 14 Other DSURs. 15 Lack of efficacy.
+16 Region-specific information. 17 Late-breaking information. 18 Overall safety assessment (18.1
+risks, 18.2 benefit-risk considerations). 19 Summary of important risks. 20 Conclusions.
+
+Notes: the reporting period is normally twelve months from the development international birth
+date anniversary; state the actual dates supplied. Keep sections that do not apply and say why.

pharmaskills/_skills/aggregate-report/references/pbrer-sections.md

@@ -0,0 +1,19 @@
+# PBRER section order (ICH E2C(R2))
+
+Notes written for this project from the public table of contents; no guideline text copied.
+Machine-readable list: `src/pharmaskills/standards/data/pbrer.yaml` (UNSURE: wording of some
+subsection titles; verify against the published PDF before quoting).
+
+1 Introduction. 2 Worldwide marketing authorisation status. 3 Actions taken in the interval for
+safety reasons. 4 Changes to reference safety information. 5 Estimated exposure (5.1 clinical
+trials, 5.2 marketing). 6 Summary tabulations (6.1 reference information, 6.2 clinical trial
+serious events, 6.3 post-marketing). 7 Significant clinical trial findings (7.1 to 7.5). 8
+Non-interventional studies. 9 Other clinical trials and sources. 10 Non-clinical data. 11
+Literature. 12 Other periodic reports. 13 Lack of efficacy in controlled trials. 14
+Late-breaking information. 15 Signal overview. 16 Signal and risk evaluation (16.1 to 16.5).
+17 Benefit evaluation (17.1 to 17.3). 18 Integrated benefit-risk analysis (18.1 context, 18.2
+evaluation). 19 Conclusions and actions. 20 Appendices.
+
+Notes: the PBRER is cumulative in its evaluation, with interval data presented separately where
+the format asks. State the interval, the IBD and the data lock point at the front. The reference
+safety information version in force at the data lock point is named in section 4.

pharmaskills/_skills/aggregate-report/scripts/check_aggregate.sh

@@ -0,0 +1,8 @@
+#!/bin/sh
+# Thin wrapper: section check of a PBRER or DSUR draft.
+# Usage: check_aggregate.sh <file.docx|file.md> <pbrer|dsur> [report.json]
+set -eu
+doc="${1:?usage: check_aggregate.sh <file> <pbrer|dsur> [report.json]}"
+std="${2:?usage: check_aggregate.sh <file> <pbrer|dsur> [report.json]}"
+report="${3:-${std}_structure_report.json}"
+exec pharmaskills docs check-structure "$doc" --standard "$std" --out "$report"

pharmaskills/_skills/aggregate-report/templates/dsur-outline.md

@@ -0,0 +1,44 @@
+# Development Safety Update Report: [DATA NEEDED: product]
+
+Reporting period: [DATA NEEDED: start] to [DATA NEEDED: end]. Region assumed: [DATA NEEDED].
+Development international birth date: [DATA NEEDED: DIBD]. Data lock point: [DATA NEEDED].
+
+## 1. Introduction
+## 2. Worldwide marketing approval status
+## 3. Actions taken in the reporting period for safety reasons
+## 4. Changes to reference safety information
+## 5. Inventory of clinical trials ongoing and completed during the reporting period
+## 6. Estimated cumulative exposure
+### 6.1 Cumulative subject exposure in the development programme
+### 6.2 Patient exposure from marketing experience
+## 7. Data in line listings and summary tabulations
+### 7.1 Reference information
+### 7.2 Line listings of serious adverse reactions during the reporting period
+### 7.3 Cumulative summary tabulations of serious adverse events
+## 8. Significant findings from clinical trials during the reporting period
+### 8.1 Completed clinical trials
+### 8.2 Ongoing clinical trials
+### 8.3 Long-term follow-up
+### 8.4 Other therapeutic use of investigational drug
+### 8.5 New safety data related to combination therapies
+## 9. Safety findings from non-interventional studies
+## 10. Other clinical trial or study safety information
+## 11. Safety findings from marketing experience
+## 12. Non-clinical data
+## 13. Literature
+## 14. Other DSURs
+## 15. Lack of efficacy
+## 16. Region-specific information
+## 17. Late-breaking information
+## 18. Overall safety assessment
+### 18.1 Evaluation of risks
+### 18.2 Benefit-risk considerations
+
+| Dimension | Evidence | Source | Uncertainty | Assessment |
+| --- | --- | --- | --- | --- |
+| Key benefits | [DATA NEEDED] | [DATA NEEDED] | [DATA NEEDED] | [DATA NEEDED] |
+| Key risks | [DATA NEEDED] | [DATA NEEDED] | [DATA NEEDED] | [DATA NEEDED] |
+| Overall conclusion | [DATA NEEDED: qualified person] | | | |
+
+## 19. Summary of important risks
+## 20. Conclusions

pharmaskills/_skills/aggregate-report/templates/pbrer-outline.md

@@ -0,0 +1,54 @@
+# Periodic Benefit-Risk Evaluation Report: [DATA NEEDED: product]
+
+Reporting interval: [DATA NEEDED: start] to [DATA NEEDED: end] (data lock point [DATA NEEDED]).
+Region assumed: [DATA NEEDED]. International birth date: [DATA NEEDED: IBD].
+
+## 1. Introduction
+## 2. Worldwide marketing authorisation status
+## 3. Actions taken in the reporting interval for safety reasons
+## 4. Changes to reference safety information
+## 5. Estimated exposure and use patterns
+### 5.1 Cumulative subject exposure in clinical trials
+### 5.2 Cumulative and interval patient exposure from marketing experience
+## 6. Data in summary tabulations
+### 6.1 Reference information
+### 6.2 Cumulative summary tabulations of serious adverse events from clinical trials
+### 6.3 Cumulative and interval summary tabulations from post-marketing data sources
+## 7. Summaries of significant findings from clinical trials during the reporting interval
+### 7.1 Completed clinical trials
+### 7.2 Ongoing clinical trials
+### 7.3 Long-term follow-up
+### 7.4 Other therapeutic use of medicinal product
+### 7.5 New safety data related to fixed combination therapies
+## 8. Findings from non-interventional studies
+## 9. Information from other clinical trials and sources
+## 10. Non-clinical data
+## 11. Literature
+## 12. Other periodic reports
+## 13. Lack of efficacy in controlled clinical trials
+## 14. Late-breaking information
+## 15. Overview of signals: new, ongoing or closed
+## 16. Signal and risk evaluation
+### 16.1 Summary of safety concerns
+### 16.2 Signal evaluation
+### 16.3 Evaluation of risks and new information
+### 16.4 Characterisation of risks
+### 16.5 Effectiveness of risk minimisation
+## 17. Benefit evaluation
+### 17.1 Important baseline efficacy and effectiveness information
+### 17.2 Newly identified information on efficacy and effectiveness
+### 17.3 Characterisation of benefits
+## 18. Integrated benefit-risk analysis for authorised indications
+### 18.1 Benefit-risk context
+### 18.2 Benefit-risk analysis evaluation
+
+| Dimension | Evidence | Source | Uncertainty | Assessment |
+| --- | --- | --- | --- | --- |
+| Condition and unmet need | [DATA NEEDED] | [DATA NEEDED] | [DATA NEEDED] | [DATA NEEDED] |
+| Key benefits | [DATA NEEDED] | [DATA NEEDED] | [DATA NEEDED] | [DATA NEEDED] |
+| Key risks | [DATA NEEDED] | [DATA NEEDED] | [DATA NEEDED] | [DATA NEEDED] |
+| Risk management | [DATA NEEDED] | [DATA NEEDED] | [DATA NEEDED] | [DATA NEEDED] |
+| Overall conclusion | [DATA NEEDED: qualified person] | | | |
+
+## 19. Conclusions and actions
+## 20. Appendices to the PBRER