ai-analyst-lab Trace one analysis
First invoke the trace skill so the machine-readable provenance record,
lightweight receipt, and shareable HTML exist. Then review the claim through
these records:
- The output states the claim.
- The receipt identifies the source, data snapshot, query, and artifacts.
- The trace records how the run unfolded.
- Logs record specific system and query events.
- The query and result provide execution evidence.
Use helpers.evals.trace.inspect_receipt on
working/trace_receipt_<analysis_id>.json to identify missing fields and
initial query risks. Then inspect the analytical details that code cannot infer
from syntax alone: population, filters, time window, timezone, exclusions,
grain, join cardinality, missingness, referential integrity, freshness,
reconciliation, and whether the claim stays within the evidence.
For a query with a join and aggregate, test the grain directly. Compare the candidate total with a defensible reference query when one exists. A query that runs successfully may still answer the wrong question.
Save the written review beside the analysis. Report the review path and the HTML trace path together. Report missing evidence as missing. A complete receipt makes an analysis inspectable. It does not make the analysis correct.