Tell, Don't Cut

What Dad Does Not See

Person marks catch who is in a picture. A second kind of leak is what is in it: a delivery label on a box at the edge of the frame, a notification sliding down a phone on the desk, an address on an envelope, a name on a screen. People are bad at noticing these in their own pictures, because they look at the subject and their eyes skip the clutter they live with. Dad named this as the second rule on the same day as the first:

아빠가 못 보는 걸 피파가 볼 때가 많아... 그럴 땐 아빠한테 알려줘. — 아빠 (2026-11-02)

Pippa often sees what Dad does not; when she does, she should tell him. Note the verb. Tell, not cut.

Reading a Frame Before It Leaves

Every frame that goes out through the export door is read first. The door fits the frame to a 2560-pixel view and runs the same native Vision helper, this time for its on-screen text. Each recognized line is matched against a short list of patterns, in order:

registration numbers, card numbers, phone numbers, account numbers;
e-mail addresses;
addresses (Korean administrative suffixes, apartment unit numbers, or a number and an English street name);
delivery-label words, messenger and mail notifications;
names with an honorific (Korean -님, or Mr, Mrs, Ms, Dr);
and, separately, a readable document: twelve or more lines of text in one frame means a page someone could read, whatever it says.

Two small rules make the findings useful rather than noisy. One stretch of text is one finding (a phone number is not also reported as an account number). And every finding carries where in the frame it is, as one of nine areas ("top left", "bottom centre") plus its box, so a person can find it in a second.

Why Not Blur It Automatically?

The obvious next step would be to blur every finding before the frame leaves. Oriel deliberately does not, and the reasons are worth taking apart:

Patterns are wrong in both directions. A game's score can look like a phone number; a real address can be split across two recognized lines and match nothing. An automatic blur would deface innocent frames and still miss real leaks, while everyone believed the problem was handled.
The decision has context the code lacks. Dad's own channel name on a screen is not a leak; a neighbour's name on a parcel is. Only the person who owns the picture knows which.
A silent edit is a lie about the picture. It is the same spirit as the family's rule against silent truncation: never cut quietly; refuse or report out loud.

So the findings return with the export, go into the activity log, and the export action's answer tells Pippa, in so many words, to tell Dad before the frame goes anywhere ("Tell Dad before it goes anywhere: phone number "…" (top left)"). Nothing is cut. Dad decides.

The Same Habit Everywhere

The rule is not limited to exports. Whenever Pippa looks at something that may go out (a screenshot for a post, a frame for a thumbnail, the office picture on a broadcast), she watches for details Dad missed and names where they are. The export door is where that habit became code; the habit itself is older and wider.

Code

Patterns, one finding per stretch, and where it is (trimmed excerpt, backend/marks.py)·python

SENSITIVE = [
    ("registration number", re.compile(r"\b\d{6}\s?-\s?[1-4]\d{6}\b")),
    ("card number", re.compile(r"\b(?:\d{4}[ -]){3}\d{4}\b")),
    ("phone number", re.compile(r"(?:\+82[ -]?|\b0)1[016789][ -.]?\d{3,4}[ -.]?\d{4}\b|\b0\d{1,2}-\d{3,4}-\d{4}\b")),
    ("account number", re.compile(r"\b\d{2,6}(?:-\d{2,7}){2,}\b")),
    ("e-mail", re.compile(r"\b[\w.+-]+@[\w-]+\.[\w.]+\b")),
    ...  # address, delivery label, notification, name with honorific
]
DOCUMENT_LINES = 12


def where(box) -> str:
    cx, cy = box["x"] + box["w"] / 2, box["y"] + box["h"] / 2
    row = "top" if cy < 1 / 3 else "bottom" if cy > 2 / 3 else "middle"
    col = "left" if cx < 1 / 3 else "right" if cx > 2 / 3 else "centre"
    return f"{row} {col}"


def sensitive(result) -> list:
    """What a frame's on-screen text would give away, with where in the frame."""
    findings = []
    lines = result.get("text") or []
    for line in lines:
        text = str(line.get("text") or "")
        taken = []
        for kind, pattern in SENSITIVE:
            match = pattern.search(text)
            # One stretch of text is one finding: a phone number is not also an account.
            if match and not any(match.start() < end and start < match.end() for start, end in taken):
                taken.append(match.span())
                findings.append({"kind": kind, "text": match.group(0)[:80], "where": where(line["box"])})
    if len(lines) >= DOCUMENT_LINES:
        findings.append({"kind": "readable document", "where": "whole frame"})
    return findings

Run the same idea on recognized lines of your own (runnable; every value here is made up)·python

import re

PATTERNS = [
    ("phone number", re.compile(r"\b0\d{1,2}-\d{3,4}-\d{4}\b")),
    ("e-mail", re.compile(r"\b[\w.+-]+@[\w-]+\.[\w.]+\b")),
    ("delivery label", re.compile(r"tracking|ship to", re.I)),
]


def where(x, y):
    row = "top" if y < 1 / 3 else "bottom" if y > 2 / 3 else "middle"
    col = "left" if x < 1 / 3 else "right" if x > 2 / 3 else "centre"
    return f"{row} {col}"


lines = [  # (text, centre x, centre y) as a text recognizer might report them
    ("LEVEL 12  SCORE 010-1234-5678", 0.2, 0.1),     # a game HUD that looks like a phone number
    ("someone@example.com", 0.8, 0.9),
    ("Tracking no. 0000-0000", 0.5, 0.95),
]
for text, x, y in lines:
    for kind, pattern in PATTERNS:
        if pattern.search(text):
            print(f"tell Dad: {kind} at {where(x, y)}: {text!r}")

Exercise

Take three screenshots of your own desktop at an ordinary moment (not staged), run the Swift OCR script from track four on each to get recognized lines, and feed them to the pattern sketch. Write down every finding and every miss you spot by eye. Then decide, for each finding, whether you would have wanted it blurred automatically.

Hint

You will very likely find at least one false positive (a version number, a score, a date) and one thing your eye catches that no pattern does (a name without an honorific, a partial address). Both are the argument for telling instead of cutting.