Redact Round-Trip

Overview

What Redact Round-Trip is

It lets you get help from an approved AI assistant on real work — examination notes, loan tapes, evidence — without exposing names, account numbers, or other identifiers.

You paste your text in. The tool finds the sensitive parts and replaces each one with a neutral placeholder (a token) such as [[INSTITUTION_1]]. You copy that safe version into your assistant and ask your question. When the assistant answers — using the same tokens — you paste the answer back, and the tool swaps the real names in again. You get a finished draft with the real details restored, and the sensitive data never left your computer.

The core idea

How it works: the round-trip

Everything the tool does is one loop. There are two halves — masking on the way out, and restoring on the way back.

  1. You paste your real notes or table
            │
  2. The tool highlights what it found  →  you review and adjust
            │
  3. Copy the masked version (names → tokens)
            │
  4. Paste it into your approved assistant, ask your question
            │
  5. Paste the assistant's reply back
            │
  6. The tool puts the real values back → finished draft

Steps 1–3 happen on the Mask outgoing tab. Steps 5–6 happen on the Restore reply tab. The tool quietly keeps a private list (the session map) of which token stands for which real value, so it can reverse everything in step 6.

Plain-language concepts

Six things worth understanding

1. Tokens

A token is a stand-in label that replaces a real value, like [[PERSON_1]] for a person's name. Tokens are consistent: the same real value always becomes the same token, everywhere it appears. So if the assistant reasons about [[INSTITUTION_1]], that logic still makes sense — and restores correctly — wherever it shows up.

2. Mask, keep, or offset

Every detected item is handled in one of three ways:

By default, identifiers are masked and analytical figures are kept — because masking a dollar amount or a ratio would make the assistant's analysis useless. Offset is an optional extra for the most cautious users (covered later).

3. Free text vs. tables

The tool recognises two kinds of input and handles each correctly. Free text is ordinary writing — notes, memos, narratives. Tables are rows and columns — a CSV file, a range copied from Excel, or JSON records. When you paste a table, the tool works column by column instead of word by word, which is faster and more accurate. It tells you which kind it detected, and you can switch with one click.

4. The session map

This is the private list of token ↔ real value pairs for your current session. It lives only in your computer's memory and is what makes restoring possible. It is cleared when you press Clear session or close the tab.

5. Assistive, not automatic

Automatic detection is helpful but never perfect — it can miss an unusual name or an oblique reference. That's why the tool requires you to review before anything can be copied. Think of it as a fast first pass that you confirm, not a guarantee you can skip.

6. Local-only

All of the work happens inside this one page on your device. The tool makes no network calls — your text is never sent anywhere. You can confirm this yourself (see Privacy & security).

Get going in 90 seconds

Quick start

1. Paste your text into the box on the Mask outgoing tab. The tool detects identifiers automatically.
2. Review the highlights. Masked items are gold; kept items are grey. Click any item to switch it. Missed something? Select it and press Mask selection.
3. Tick the confirmation box — "I have reviewed the masked output…". This unlocks the copy button.
4. Copy the masked output and paste it into your approved assistant. Ask your question as normal.
5. Restore: go to the Restore reply tab, paste the assistant's answer, and press Restore original values. You get your finished draft with the real names back in.

Use case

Examination notes (free text)

The everyday case: you have written notes and want the assistant to summarise, critique, or draft from them.

First Mercantile Bank — Q2 2026 exam.
CEO J. Reynolds. Classified assets $4.2M
(Q1 $3.1M); CRE concentration at 220%.
Routing 021000021; borrower SSN
123-45-6789.

[[INSTITUTION_1]] — Q2 2026 exam.
CEO [[PERSON_1]]. Classified assets $4.2M
(Q1 $3.1M); CRE concentration at 220%.
Routing [[ROUTING_1]]; borrower SSN
[[SSN_1]].

Notice that $4.2M, $3.1M, 220% and Q2 2026 were kept — the assistant needs them to reason. Only the identifiers were masked.

Use case

A table or loan tape (CSV or Excel)

Paste a CSV file, or copy a range straight out of a spreadsheet (Excel copies as tab-separated — the tool handles both). It reads the columns and proposes one rule per column: mask, keep, or scan per cell.

Loan ID,Borrower,SSN,Balance,Rate,Notes
44219-A,Acme Holdings LLC,123-45-6789,$1250000,5.25%,Guarantor J. Reynolds
44220-B,Acme Holdings LLC,987-65-4321,$430000,6.10%,CRE flagged

Loan ID,Borrower,SSN,Balance,Rate,Notes
[[LOAN_1]],[[PERSON_1]],[[SSN_1]],$1250000,5.25%,Guarantor [[PERSON_2]]
[[LOAN_2]],[[PERSON_1]],[[SSN_2]],$430000,6.10%,CRE flagged

Three things to see here:

• The table shape is preserved — same columns, same rows, same header — so you can paste the answer back into your spreadsheet.
• Balance and Rate were kept; the Borrower and SSN columns were masked.
• "Acme Holdings LLC" became [[PERSON_1]] in both rows. Identical values share one token, so the assistant can still group and total by borrower — and it all restores correctly.

Use case

JSON records

If you paste a list of JSON records, the tool treats it as a table (each field is a column) and gives it back as valid JSON — with numbers and true/false values keeping their type.

[
  {"name":"Jane Roe","ssn":"123-45-6789","bal":1000},
  {"name":"Jane Roe","ssn":"555-66-7777","bal":2000}
]

[
  {"name":"[[PERSON_1]]","ssn":"[[SSN_1]]","bal":1000},
  {"name":"[[PERSON_1]]","ssn":"[[SSN_2]]","bal":2000}
]

bal stayed a number (1000, not "1000"), and the repeated name shares one token — same rules as any other table.

Use case

Restoring the assistant's reply

This is the second half of the round-trip. The assistant answers using the tokens; you bring the real values back.

[[INSTITUTION_1]] shows rising
classified assets. [[PERSON_1]]
should address the CRE concentration.

First Mercantile Bank shows rising
classified assets. J. Reynolds
should address the CRE concentration.

1. Open the Restore reply tab.
2. Paste the assistant's response (the one that still has the tokens in it).
3. Press Restore original values. The tool reports how many it restored and flags anything unusual (see Restore messages).

Use case

Catching something the detector missed

Automatic detection won't catch every oblique reference — for example, "the developer with the 220% concentration" names no one but identifies someone. When you spot a miss, mask it by hand:

1. In the paste box, select the text you want to hide.
2. Press Mask selection. It becomes a token like any other detected item, and restores the same way.

Power feature · Advanced

Custom word lists: always-mask & never-mask

Switch the tool to Advanced (top right) to open Customization & profiles. Two lists let you teach it your world:

• Always mask — terms that must always be hidden, even if the detector wouldn't catch them: your institution's name, a counterparty, an examiner roster. One per line; you can add a type, e.g. First Mercantile :: INSTITUTION.
• Never mask — terms you don't want flagged, like a public regulator's name. One per line. These are left untouched.

Greenfield Co was cited by the
Federal Reserve for the LF-778899 gap.

[[INSTITUTION_1]] was cited by the
Federal Reserve for the [[LOAN_1]] gap.

"Greenfield Co" was masked because you told it to; "Federal Reserve" was left alone; and the LF-778899 code was caught by a custom pattern (next).

Power feature · Advanced

Custom patterns

If your institution uses its own code format — say loan files look like LF-778899 — add a pattern so the tool always catches it. Patterns use standard "regular expression" syntax, with an optional type:

LF-\d{6} => LOAN_FILE

As you type a pattern, a live tester shows how many matches it finds in your current text (or flags an invalid pattern), so you can get it right before applying.

Power feature · Advanced

Token style & labels

You can change how tokens look, so the masked text reads naturally for your team:

• Wrapper — choose [[ TOKEN ]] (default), ⟦ TOKEN ⟧, or {{ TOKEN }}.
• Labels — rename a type's stem, one per line: INSTITUTION=BANK turns [[INSTITUTION_1]] into ⟦BANK_1⟧.

⟦BANK_1⟧ — Q2 2026 exam. CEO ⟦PERSON_1⟧.

Restoring works exactly the same whichever style you pick — the tool reverses your chosen wrapper automatically.

Power feature · Advanced

Offset amounts & shift dates

For the most cautious users who don't want even real numbers leaving the device, Transform modes change figures by a consistent amount that's reversed automatically when you restore:

• Offset dollar amounts by a factor (e.g. ×1.07). Every amount scales the same way, so ratios and relative sizes are preserved — the assistant's analysis stays valid.
• Shift dates by a number of days (e.g. +37). Every date moves the same way, so intervals and aging stay correct.

Balance $1,250,000
Originated 2024-03-12
Rate 5.25%

Balance $1,337,500
Originated 2024-04-18
Rate 5.25%

Power feature · Advanced

Profiles — set once, share with the team

A profile bundles everything you've configured — word lists, patterns, token style, transform settings — into one file. It contains rules only, never any of your content.

• Export profile saves it as a small file you can send to colleagues.
• Import loads a profile someone shared, configuring the tool instantly.
• Starter packs let you load a ready-made set of rules — Finance (the default) and a Healthcare example are included.
• Reset to defaults clears your customizations.

Power feature · Advanced

Remembered column rules

If you regularly paste tapes with the same column headers, set the column rules once and press Remember these column rules. Next time a table with matching headers comes in, those rules apply automatically — a column named SSN masks, Balance stays, and so on. The rules are saved in your profile (no content), and you can clear them anytime.

Reference

A tour of the screen

Simple vs. Advanced

The switch at the top right toggles between two views. Simple is the everyday five-step flow with nothing extra. Advanced adds the customization tools and a diagnostics panel. You never need Advanced to do the core job.

The review colours

Click any highlighted item to switch how it's handled. A dashed underline on a detected item means lower confidence — worth a closer look.

The session map

The panel on the right lists every token and the real value it stands for. Use it to sanity-check before you copy.

The confirmation gate

The copy button stays locked until you tick "I have reviewed the masked output…". This is deliberate — it makes the human check unskippable.

Diagnostics (Advanced)

Shows how fast detection ran, whether it ran in the background, how the tool read your input, and a live count of network calls — which stays at 0.

Reference

Reading the restore messages

After you restore a reply, the tool shows one or more banners. Here's what each means:

• "N tokens restored" (green) — real values were swapped back in. All good.
• "N offset values reversed" (green) — offset amounts/dates were returned to their real figures. Reminds you that calculated numbers stay in the offset domain.
• "Unknown tokens in the reply" (red) — the reply contains a token that isn't in your session map. The assistant may have invented it. Verify these by hand before using the draft.
• "Not referenced by the assistant" (grey) — some of your tokens didn't appear in the reply. This is normal — the assistant simply didn't need them (e.g. a routing number). Nothing to fix.
• "No tokens were found" (grey) — the reply had no tokens. Check you pasted the assistant's tokenised answer.

Please read

What this tool does — and does not — do

Detection is assistive, not a guarantee. The tool is a fast first pass that you confirm. It can miss things — especially oblique references and institution-specific codes. The review step exists for exactly this reason, and it is required, not optional. You are responsible for what leaves the tool.

Reference

Privacy & security

• Nothing is transmitted. All detection, masking, and restoring happen on your device. The tool makes no network calls.
• You can verify it. In Advanced, the diagnostics show a live network-call count that stays at 0. You can also open your browser's network tab, or disconnect from the internet entirely — the tool keeps working.
• The session map is sensitive and temporary. It holds your real values and lives only in memory. Clear session, or closing the tab, wipes it.
• Clipboard hygiene. After you copy, the clipboard can auto-clear after 90 seconds (a setting in Advanced). The clipboard only ever holds masked text or your restored draft.
• Profiles carry no content. An exported profile contains your rules only — never names, numbers, or any text you've masked. It's safe to share.

Reference

Frequently asked questions

Reference

Troubleshooting

• My table was read as free text (or vice-versa). Use the link in the banner under the paste box — "Treat as a table instead" / "Treat as free text" — to switch.
• The copy button is greyed out. Tick the confirmation box first. It also requires at least one masked item.
• A restored draft still has a token in it. That token wasn't in your session map — likely the assistant invented or altered it. The tool flags these; verify by hand.
• Restore did nothing. Make sure you masked something first (the session map must exist), and that you pasted the assistant's tokenised reply.
• A custom pattern isn't matching. Watch the live tester as you type it — it shows matches or flags an invalid pattern.
• It feels slow on a huge file. Very large tables show a preview of the first rows while masking the whole thing on copy; for very large pastes, auto-detect pauses and you press the Detect button instead.

Reference

Glossary

Token

A neutral placeholder that replaces a real value, e.g. [[PERSON_1]]. The same value always gets the same token.

Mask

To replace a sensitive value with a token.

Keep

To leave a value unchanged — used for amounts, ratios, and dates the assistant needs.

Offset / shift

An optional transform that changes amounts or dates consistently and reverses on restore.

Session map

The private, in-memory list pairing each token with its real value. Cleared on close or "Clear session".

Round-trip

The full loop: mask on the way out, restore on the way back.

Free text vs. table

Two input kinds the tool handles differently — ordinary writing vs. rows and columns.

Per-cell

A column rule that scans each cell of a free-text column (like "Notes") for identifiers.

Profile

A shareable bundle of your rules and settings — content-free.

CSI

Confidential supervisory information. Some policies restrict its external use even when de-identified.