Pattern scout

Scan data systematically and return candidate patterns the user may find striking. The user judges interest; you report evidence strength. Run only when asked.

1. Protect the user's own view

Before showing candidates, check whether the user has recorded their own observations for this data (ledger observations.md). If not, ask whether they want to note them first, so the scout list does not anchor them. Respect a "just run it".

2. Define the scan before running it

Write down, and log in the ledger (analyses.md):

• the question or context (from the ledger index)
• variables and relationships to scan (e.g. all metric pairs, metric vs condition, per-group differences)
• the number of comparisons this implies
• thresholds and methods

Defining the space first is what makes the multiple-testing correction honest.

3. Scan

For each candidate:

• effect size with CI (bootstrap where appropriate)
• effective n
• robustness: rank-based version, leave-one-group-out (target, batch, family), outlier removal
• multiple-testing correction across the whole scan (e.g. Benjamini–Hochberg q-value), or explicit "exploratory, uncorrected"
• artefact flags: driven by one group or outlier; confounded with batch, length, date; ceiling/floor; trivially explained by construction (e.g. a composite score correlating with its own component)

4. Report candidates

• Rank by evidence strength (effect size, robustness, corrected significance), not by how interesting you think it is.
• Show the top 5–10 in a table with caption per reporting-numbers; write the full list to a file and point to it.
• IDs: S1, S2, … (scout candidates). They are not observations until the user adopts them.

Then one line: total comparisons made, how many survive correction, and what fraction would be expected by chance.

Familiarity flag (see co-scientist) only if a candidate is likely to be called novel and the concept is specifically familiar.

5. Hand over

• The user picks which candidates are interesting. Adopted ones become observations (O-ids) in the ledger, marked recorded by: scout.
• Never phrase candidates as conclusions or mechanisms. "X and Y co-vary (ρ = 0.6)", not "X drives Y".
• Log the scan, including null results and the comparison count, in analyses.md.