Create Data Inventory
Inspect the task's authorized data locations before formal analysis, modelling, training, or evaluation. Use bounded metadata and representative-content checks to identify every task-relevant dataset, subject, session, run, partition, and label source that is visible within the assigned scope. Reconcile counts across filenames, metadata, labels, and loaded shapes; state unresolved discrepancies instead of guessing.
Write one canonical Markdown document at docs/specs/data-inventory.md. Create
the parent directory when needed. If the document already exists, read it first
and revise it in place so downstream agents retain one authoritative path.
Include:
- the task scope and every location inspected;
- a table of discovered data sources with paths, formats, sizes or record counts, subjects, sessions, runs, trials or other natural grouping units;
- loaded structure where relevant: shapes, axes, dtypes, value domains, feature ordering, labels and label mappings;
- documented or observed split relationships and grouping boundaries;
- reconciled totals and the checks used to obtain them;
- missing metadata, ambiguous mappings, unreadable inputs, discrepancies, and any parts of the assigned scope that could not be inspected;
- concise implications that Experimentalist, Librarian, Engineer, and Auditor must preserve when designing or reviewing later work.
Keep the inventory descriptive. Do not train models, choose methods, freeze hyperparameters, or treat inventory inspection as empirical validation. Link supporting metadata or inspection outputs when they are useful, but keep the inventory self-contained enough that another agent can use it without repeating the discovery work.
Before handoff, verify that every discovered task-relevant grouping is accounted
for, totals agree or are explicitly unresolved, and every referenced path exists.
Return docs/specs/data-inventory.md as the primary artifact and direct dependent
agents to read it before using the data.
