DISCOVERY ENGINEDOCS

Stand-ins and what the AI sees

How the app detects personal details and gives them stand-ins like PERSON A, how the Show stand-ins switch works, and exactly what your AI agent sees in this version.

#What a stand-in is

While it reads a case, the app finds personal details in the text, on your computer:

  • people's names and organizations,
  • account and card numbers,
  • addresses, phone numbers and e-mail addresses,
  • government ID numbers and dates of birth.

Each one gets a stand-in: a neutral label such as PERSON A, ORG 3, ACCOUNT 1 or ADDRESS 2. One real person gets one stand-in across the whole case, across every spelling the app has matched to them, so Daniel Reeves, D. Reeves and dreeves@… can all be PERSON A.

Dates and money amounts are not replaced. The evidence depends on them ("On March 3 … $48,000").

This detection runs in the Find personal details and Link names stages you see on Progress. It is also what lets search find every spelling of a name, and what the People, Graph and Connections screens are built on.

#What you see on screen

Real names. The app is your private research tool, so it shows the documents as they are written.

Hover over a detected name and a tooltip shows its stand-in, for example Stand-in: PERSON A. That's a quick way to see which spellings the app has tied to the same person.

#The Show stand-ins switch

At the bottom of the left column is a switch: Show stand-ins. Turn it on and every screen shows what a sanitized view looks like:

  • names, accounts and addresses in text appear as their stand-ins;
  • the tooltips that would show real values are gone;
  • document titles and file names are hidden (they can hold names), and documents are shown by ID instead;
  • on page images in the document viewer, detected names are covered with their stand-ins. If the covers can't be drawn for a page, the page isn't shown at all rather than shown uncovered.

Use it for screenshots, screen sharing and demos, or simply to see exactly what the sanitized version of a passage looks like. The switch is off each time the app starts.

#What your AI agent sees in this version

This was a deliberate decision for this version, so the agent can name files and people plainly and link straight to the originals. The app says so in three places:

  • Setup and Settings → This case → Let AI agents use this case: "…with real names and file names. Every call is logged."
  • The Start Up screen, under the prompt: the agent reads the case "as it is written, with real names and file names".
  • The confirmation when you turn agent access on: "It sees the case as written, real names and file names included… Whatever it reads may be sent to the agent's own AI provider."

What still holds:

  • The app itself sends nothing anywhere. Reading, detection, search, the graph and exports all run on your computer. The app's own built-in cloud AI is switched off in this version.
  • The prompt names cases by ID only. The prompt you paste into your agent doesn't contain case names or folder names, and writes your home folder as ~. If one of your chosen folders has a name the app found in the case files, the prompt is withheld and the Start Up screen asks you to choose folders whose names aren't people's names.
  • You decide whether an agent can use a case at all. Turn Let AI agents use this case off in the case's Settings and every agent command on that case is refused.
  • Every agent command is logged on the Agent log screen.

The engine also has a stand-in mode for agents, in which the command-line tool only ever shows stand-ins. It is off by default in this version and is not a setting in the app; it can only be turned on in a special build (an advanced, build-time option). See Agent access and privacy.

#Stand-ins in exports

Exports for the attorney use real names by default. Switch Use real names off on the Exports screen to get a sanitized copy instead:

  • names, accounts and addresses appear as stand-ins;
  • PDFs and pictures become images with the personal details blacked out, with no text left under the boxes;
  • the exhibit index uses sanitized titles and shows document IDs in place of file paths;
  • a document whose personal details couldn't be blacked out with certainty is left out of the sanitized copy, with the reason listed.

See Exports.

#How good is the detection?

Name detection is automatic and English-focused. On the fictional test case it finds the planted names reliably, but no detector is perfect:

  • It can miss a name written in an unusual way (an OCR slip, a nickname).
  • It can mistake an ordinary word for a name.

Merging spellings on the People screen improves both search and the graph. For anything that will leave your hands in sanitized form, look over the result before you send it.