trace

/trace — expose the query logic behind every number

Renders one self-contained HTML that ties each reported number (a finding) back to the query that produced it, labeled by confidence: cited (the agent named the query), value-match (a query's captured result_value equals the number), or inferred (nearest query in time). Unmatched findings and orphan queries are shown, not hidden — an unverified number is the most important thing to surface. This is the on-demand artifact for any "prove it" moment.

It reads the provenance infrastructure: the query log (hook-stamped with analysis_id + result_value), the findings manifest, and the reconciler.

Steps

1. Resolve the analysis. Read the current analysis_id and active dataset:

   python3 -c "
   import sys; sys.path.insert(0, '.')
   from helpers.knowledge.analysis_context import current_analysis_id
   from helpers.provenance.eval_driver import _active_dataset
   print(current_analysis_id(create=False) or '', _active_dataset())
   "
   

   If there is no current analysis, there is nothing to trace yet — say so and stop (or, for a past run, point build_trace at that analysis_id explicitly).

2. Build + render the trace. Date is today (date '+%Y-%m-%d'):

   python3 -c "
   import sys; sys.path.insert(0, '.')
   from helpers.provenance.trace_viewer import build_trace
   print(build_trace('<analysis_id>', '<dataset>', '<YYYY-MM-DD>'))
   "
   

   This reconciles (writes working/provenance_<analysis_id>.json) and renders working/trace_<analysis_id>.html. Both are gitignored working files.

3. Open it. open working/trace_<analysis_id>.html (macOS). It's self-contained and projection-friendly — large type, collapsible SQL, colored confidence badges.

4. Read it out. Walk the findings top to bottom: the number, its badge, the SQL. Call out anything unmatched (a number with no query behind it) — that's the honesty check, and the thing to fix.