ai-analyst-lab/ai-analyst-plugin/trace

/trace: tie every number to its source

Purpose

Answer "where did that number come from?" with evidence, not memory. Every number in a brief or readout gets linked to the exact computation and source data that produced it.

When to Use

• The user asks where a number came from, or to show the work.
• Before any deliverable leaves the session (analyst-core rule: trace numbers to source).
• Reviewing an analysis produced earlier in the session or found in .knowledge/analyses/.

Instructions

1. Collect the findings. List every specific number in the deliverable being traced: headline

figures, table cells that carry the argument, chart values called out in text.

1. Collect the computations. Gather the queries and code run this session (re-read your own

steps; if the analysis logged queries to .knowledge/query-log.md, read that too).

1. Build the trace table and include it in the output (or save as trace.md next to the

deliverable if the user wants a file):

Confidence labels:

• cited: the computation for this number was run this session and is shown.
• value-match: a run computation produced this value, but the deliverable did not name it.
• inferred: no run computation produced it; state where it came from (an input doc, an

assumption) or mark it unverified.

1. Surface the gaps loudly. Unverified numbers and orphan queries (run but unused) are the

most important rows. Never hide them. If a headline number is inferred or unverified, say so at the top of the trace, not in a footnote.

1. Offer the fix. For any unverified number, offer to recompute it live from the source data

so it can be promoted to cited.

Notes

• This is a reading-and-reporting skill: no scripts, no extra infrastructure. The evidence is the

session's own work plus the files in the working folder.

• Pairs with the reconcile check (parts sum to totals) and the reliability skill (run it again).