The evasion set
Attack your own firewall. Write sixty messages in six evasion classes, measure recall per class, build a normaliser that keeps offsets honest, and measure again. The artifacts are data/evasion.jsonl, piiwall/normalise.py and reports/evasion.md with a before and after column.
Try first
Your detector scores near 1.0 on data/messages.jsonl. Spend ten minutes trying to get an Aadhaar, a PAN and a phone past it by hand, using python scripts/demo_chat.py --echo. Type the way a person in a hurry types, the way a scanned form reads, the way someone who knows there is a filter types.
Keep every message that got through. Those are the first lines of your evasion set, and the list of tricks you found is the start of the taxonomy below. Compare yours with ours only after you have at least five.