--- name: audit-dataset title: Audit a Dataset in a Child Agent description: Inspect one dataset in a child, adapt checks to its real files and the parent's question, and return grounded findings, source references, optional outputs and unresolved scientific questions. effect: spawn_subagent_with_skill keywords: - level:L0 - data-onboarding - delegation --- You are already running this skill in the delegated child. Complete the assigned inspection directly; do not recursively delegate it. The assignment must identify the source bundle and should identify the question or audit scope. If the source is unreadable, report the actual access problem. Write notes, code and outputs to the workspace so the source stays as delivered, by default in the Agent workspace state at `/datasets//` (resolved from `CLIO_AGENT_WORKSPACE_STATE_DIR`), or wherever the parent's assignment says. 1. Load `onboard-dataset` and inspect the files needed for the assignment. Load the domain or source skills the parent names for scientific and format checks, and follow a source skill's version contract when it states one. Scripts, cards and schemas are optional aids. Adapt or replace them when their assumptions do not fit. 2. Keep uncertain treatment meanings, units, calibration and identity mappings explicit. Return these questions to the parent rather than guessing or editing the source. Continue checks that do not depend on their answers. 3. Check calculations and outputs against source observations, counts and join behavior. Process completion or a passing example schema alone does not establish scientific validity. Do not convert a whole dataset merely to fill a required set of tables. The parent owns final presentation. 4. Return a concise account of inspected scope, evidence, findings, limitations, questions, paths of any notes/code/outputs, and checks with their results. For reusable code, include its invocation and dependencies. If no loader or views were needed, say so. Identify any specific unfinished work. The parent checks the evidence supporting its answer: inspecting a sample, repeating a calculation, comparing counts or rerunning selected code when warranted. A mandatory second execution of the entire pipeline is not required. A partial result must be labelled accurately. Continue until the assigned scope is satisfied or a concrete access or interpretation dependency remains.