realdealcpa-vr/iron-jarvis/pii-redaction

PII Redaction

Produce a privacy-safe copy of a document with the PII gone and everything else — format, styling, layout — intact. The redact_pii tool does the rewriting; your job is to drive it well and to catch the PII that patterns alone cannot see.

Workflow

There is also a UI for this — Documents → Redact PII — which scans, shows

every finding with a checkbox, and asks where to save. If the user would

rather tick boxes than describe what to remove, point them there. This

playbook is the conversational equivalent and owes them the SAME two steps.

1. Locate the document(s). If the user named a file, use it; otherwise

file_search / list_files in the working folder. If several files match, confirm which ones before redacting.

1. Read before you scan. read_document the source and identify the

UNSTRUCTURED PII that regex cannot catch: person names (taxpayer, spouse, dependents), employer names, usernames — anything identifying that the user would want gone. Collect the exact strings.

1. Scan and SHOW THE LIST. Call redact_scan with the path and those

strings as extra_terms. Present the numbered findings and ask which to remove — "all of them?" is a fine question, being told after the fact is not. Wait for the answer.

1. Confirm where it goes. State the destination before writing. The default

is <name>.redacted.<ext> beside the source; if the source sits outside the workspace the tool falls back to the workspace ROOT, which is rarely what anyone wants — pass an explicit output_path in that case rather than letting the file land somewhere the user has to hunt for.

1. Pick the style from the user's words.

• "black out" / "redact" → style: "black" (same-length █ blocks — the

default; layout is preserved exactly).

• "label" / "tag" → style: "label" ([SSN], [NAME]-style tags).
• "remove" / "strip" / "delete" → style: "remove".

When the user didn't specify, use black.

1. Call `redact_pii` with `terms` = the confirmed values. Passing terms

redacts EXACTLY those strings and nothing else, which is what makes the confirmation meaningful. Only fall back to auto-detection (omitting terms) when the user is not available to confirm — and say so in your report. Use categories when they asked for specific kinds ("just the SSNs" → categories: ["ssn", "itin", "ssn_labeled"]).

1. Verify. read_document the OUTPUT file and check nothing identifying

slipped through — especially name variants (initials, "Mr. Smith", possessives) and values with unusual formatting. If anything remains, run redact_pii again on the output with those strings as extra_terms (chain: give it an explicit output_path).

1. Report. Tell the user the output file name/location, the style used,

and the redaction counts by category.

Hard rules

• Never redact without showing the findings first when the user is there to

answer. Skipping redact_scan produces a file they cannot trust: they have no idea what was taken out, or what was missed.

• Never repeat detected PII in your reply. Refer to it by category and

count ("3 SSNs, 2 addresses"), never by value. The numbered scan list is the one exception — that IS the review, and it is what they are approving.

• Never touch the original. The tool writes a new <name>.redacted.<ext>

file; if the user asks to overwrite the original, decline and point at the redacted copy instead — an irreversible PII wipe of the only copy is not recoverable.

• PDFs are rebuilt, not painted over. Tell the user the layout of a

redacted PDF is approximate but the PII content is truly removed (a cosmetic black box would leave the text extractable — that is a fake redaction and this tool refuses to do it).

• Multiple documents: redact each in turn and summarize per file.