The Numerical Obfuscation wizard

Guide 1 introduced Numerical Obfuscation in passing: a way to hide numbers — currency, percentages, counts — before text reaches a model, then restore them afterwards. This guide is the deep dive. It walks the seven steps of the wizard, explains every strategy in plain terms, and shows how to prove that your numbers come back intact before you trust the profile with real work.

What the wizard produces

The wizard builds a reusable Obfuscation Profile. You create it once, test it, save it, and then attach it to any project. From then on, the profile hides the numbers on the way to the model and restores the originals on the way back.

Opening the wizard

Go to the Privacy tab, open Obfuscation Profiles, and choose Create New Profile. The wizard runs in seven steps, shown along the top so you always know where you are. Move with Next and Back; the final step is Save Profile. You can leave at any time with Cancel.

Figure 1. The seven steps: Sample, Strategy, Configure, Tokens, Preview, Identifiers, Save. The strip across the top tracks your progress.

Step 1 — Select Sample Data and Locale

You start by choosing a sample document and a locale. The sample lets you see the strategy at work on realistic text before you commit. The locale tells the wizard how numbers are written, because that differs by language: a French figure uses a non-breaking space and a comma, a German one swaps the decimal and thousands marks, and other scripts have their own conventions entirely.

Pick the locale that matches your source — English (US), English (UK), German, French — or Mixed / Auto-detect if the document mixes formats. The sample document preview updates so you can see how numbers are recognised.

Figure 2. Step 1. Choose a sample and a locale; the preview shows how numbers are read in that format.

Step 2 — Choose an Obfuscation Strategy

A strategy is the method used to change your numbers. The wizard offers seven, shown as cards. Each is suited to a different kind of document; the next section explains them all in plain terms. For now, pick the one that fits, and the wizard moves you to its settings.

Figure 3. Step 2. The seven strategies, each with a one-line summary. Choose one to configure it.

Step 3 — Configure Strategy

Here you tune the strategy you chose. The settings change with the strategy, and each field carries a short description, so you do not need to memorise them. A few settings appear across several strategies and are worth knowing:

Setting	What it does
Scale factor	How much to multiply the numbers by. The default shrinks them to a fraction of the original.
Integers stay integers	Keeps whole numbers whole, so a count of items does not gain a decimal tail.
Deterministic (seeded)	Makes the obfuscation repeatable: the same number is hidden the same way every time, which keeps tables consistent. Leave this on.
Seed hint	An optional label (such as a project name) that fixes how the repeatable obfuscation is generated.
Placeholder prefix / pad length	For placeholder-based strategies, the shape of the tokens, for example `NUM_0001`.

Figure 4. Step 3. The settings for the chosen strategy, each with its own description.

Step 4 — Review and Refine Numeric Tokens

The wizard finds every number in the sample and lists them as tokens. For each one it shows the original text, the value it read, the kind of number it is — currency, percentage, or plain number — and the locale it used to read it. You can click a number in the document to find it in the list.

Most importantly, each token has an Include checkbox. Clear it to leave that number untouched. This is how you protect figures that must stay exactly as written, if the structural rules in step 6 do not already cover them.

Figure 5. Step 4. Every detected number, with its value, type, and locale. Clear Include to leave a number unchanged.

Step 5 — Preview Obfuscation and Restoration

This is the step that earns your trust. On the left, the Document Preview shows the result. Switch between Obfuscated, Side-by-Side, and Original to compare.

On the right, the LLM Restore Simulation proves the round trip works. Click Simulate LLM Response and the wizard generates a version of the obfuscated text with small changes added — the kind of rephrasing a real model might introduce. Then click Test Restore, and the wizard recovers your original numbers from that altered text.

Why the simulation looks slightly different

The simulated response is deliberately not identical to the obfuscated preview. It adds minor changes, such as extra spaces, on purpose. That makes it a realistic test: if your original numbers still come back after the text has been disturbed, you can trust restoration to survive a real model’s rephrasing.

Figure 6. Step 5. Side-by-side preview on the left; the simulate-then-restore test on the right that proves the originals come back.

Step 6 — Structural Identifiers

Not every number is a quantity. Section codes, standards references, and footnote markers look like numbers but must never be scaled — turning IFRS 16 into IFRS 7 would be wrong. This step keeps them verbatim.

A built-in rule already protects common identifiers such as ESRS 1, IFRS 16, IRO-2, § 5, 1.2.3, and Section 4. You can add your own under Custom identifiers — type a pattern such as R225-*, where * stands for any following characters. A Strict mode option handles references that use a non-breaking space, so a heading like Tableau 1 is protected while an ordinary phrase like les 3 années is still obfuscated.

Figure 7. Step 6. Keep section codes and standards references intact, and add your own patterns for the identifiers your documents use.

Step 7 — Save Profile and Export Artifacts

Finally, give the profile a clear name — something like Financial_Scaling_Conservative — and save it. The profile is now available to attach to any project.

If you want to inspect the result outside the app, Download Sample Artifacts saves two files: the obfuscated sample text, and the mapping that records how each number was changed. The mapping is what makes restoration possible.

Figure 8. Step 7. Name and save the profile, and optionally download the obfuscated sample and its mapping.

The seven strategies, in plain terms

Every strategy hides your numbers and can restore them through the saved mapping. They differ in how they hide them — how much of the original structure they keep, and how strong the privacy is.

Strategy	What it does	Best for
Scaling (preserve ratios)	Multiplies every number by one factor, so the relationships between them stay true.	The everyday default. Financial tables where the figures should still feel real and consistent.
Percent-Preserve	Keeps the percentage changes the same while choosing new base values.	Documents where the growth rates are the point but the absolute amounts are sensitive.
Combo: scale currency, redact percents	Scales money and plain numbers, but replaces percentages with placeholders.	Mixed financial text where the percentages are especially sensitive.
Log-Noise (multiplicative)	Adds a capped, random wobble so numbers look plausible but are not exact.	When you want believable-but-wrong figures and do not need the ratios preserved.
Bucketing / Rounding	Rounds numbers to units or turns them into ranges, reducing their precision.	When exact figures are not needed and approximate ones are safer.
Placeholders (max safety for LLM)	Replaces every number with a unique token, so the model sees no real figures at all.	The strongest privacy. Highly confidential documents.
Constrained Optimization	Changes numbers as little as possible while keeping totals, ordering, and other rules intact.	Tables that must stay internally consistent, where subtotals still have to add up.

If in doubt

For most financial work, start with Scaling — it keeps the numbers coherent. For the most sensitive documents, choose Placeholders — the model never sees a real figure. The preview step lets you compare before you decide.

Using the profile in a project

Once saved, attach the profile in the project’s privacy settings. Under Numerical Obfuscation, the Current Obfuscation Profile shows what is in use; Change… lets you pick your profile. From then on, numbers are hidden before translation and restored on export, exactly as you tested in the wizard.

Terminology used in this guide

Obfuscation Profile: A saved, reusable set of rules for hiding and restoring numbers, attached to a project.
Strategy: The method a profile uses to change numbers, such as Scaling or Placeholders.
Token: A single number the wizard has detected, with its value, type, and locale.
Mapping: The record of how each number was changed, which makes restoration possible.
Structural identifier: A code that looks like a number but is not a quantity — such as a standards reference or section number — and is kept unchanged.
Restore: Recovering the original numbers from obfuscated or model-rephrased text, tested in the Preview step.