H

Harbor Lane

Brand Guardian

Desk7
Mark-up4
Rulebook
Live channels6
Label
Press
Intake
Record
Report
Sources

Gate ran 10:34

Rulebook v7 · 4 Aug

58 rules in force · 46 of them measured or read, 6 raised by a model, 6 on a person’s checklist.

Sample data
Omni*module 04 · demoContent Studio demo
H

Harbor Lane · Guardian

Rulebook v7 · sample data

Omni*

Report

1 Sep – 13 Sep · vs last year: 1 Sep 2025 – 13 Sep 2025

TodayYesterday7 days28 daysThis monthLast month90 days12 months
Previous periodLast year

What the gate came to: how much was checked, how much was right the first time, which rules actually caught something and who sent the cleanest work. Counts of rules, never a grade for anybody's design.

Files checked

The gate

150

▼ 0.7%vs last year

Passed first time

Computed

69.3%

▲ 3.1 ptvs last year

Nothing broken on the first round

Issues per file

The gate

0.49

▲ 0.04vs last year

73 findings in all

Hours to a decision

The gate

4.4 h

▲ 0.6 hvs last year

The gate takes under a minute. The rest is a person's diary.

Sent back

The gate

46

▼ 9.8%vs last year

From outside

The gate

29

▲ 20.8%vs last year

Agency, printer and supplier files, through the same door

The period, in short

  • Files passed the first round 69 per cent of the time, against 66 per cent in the window before it3.1 pt.
  • The book did not change inside this window. Its last section, v7, arrived on 2026-08-04 and has been absorbed: the printer's specification came in as rules: ink ceiling, bleed, safety margin, embedded faces, colour space.
  • 33 per cent of everything caught came from three rules — product fills the frame the channel asks for, body type reaches 4.5:1 against its ground, nothing written on a marketplace main image. A book with a long tail and a short head is a book that is mostly teaching three things.
  • 34 of 58 rules caught nothing at all. That is not an argument for deleting them: a rule nobody breaks any more is usually a rule everybody has learned, and the year after it is deleted is the year it comes back.

Files checked

Everything that came through the gate, inside and outside, against the same length of time before it.

The gate
01020301 Sep3 Sep5 Sep7 Sep9 Sep11 Sep13 Sep
This period Last year

The rulebook did not change inside this window, so there are no hairlines on it. Widen the period to twelve months and the days the book grew a section appear on the line, each with the fortnight of failures that followed it.

Who sent it

Files through the gate in this period, by whoever put them there.

The gate
  • Content Studio

    9765%
  • In-house designer

    149%
  • The agency

    117%
  • The label printer

    117%
  • The web store

    107%
  • A packaging supplier

    75%

The three outside senders are a small share of the files and most of the round trips. An agency has the brief and the deadline and not the type licences; a printer changes artwork to make it print, which is its job and is also what the press rules are for.

Which part of the book caught it

Findings in this period, by the shelf of the rulebook the rule sits on.

Brand rulebook
  • The surface it is going to

    2838%
  • Colour

    1419%
  • Type

    1014%
  • The press

    710%
  • The words

    68%
  • The mark

    57%
  • The law, on a label

    34%

A shelf catching nothing is not a shelf doing nothing. The law section is quiet most months and it is the one where a miss is a recall rather than a reprint.

What actually caught things

Every row is a rule in the book, by its own key. Add a rule and it appears here; delete one and it goes.

The rules that fired

The twelve rules that caught the most, 56 findings between them. The share is of everything caught in the period, so these do not add up to a hundred.

The gate
RuleShelfHow it is testedCaughtSharevs before
Product fills the frame the channel asks forThe surface it is going toMeasured1115%+9
Body type reaches 4.5:1 against its groundColourMeasured710%+2
Nothing written on a marketplace main imageThe surface it is going toMeasured68%+3
Headline clear of the platform's own furnitureThe surface it is going toMeasured45%−5
Print files are CMYK, screen files are sRGBColourMeasured45%+2
Licensed for where it is goingTypeMeasured45%−2
Clear space of one H on every sideThe markMeasured45%−5
Origin named, never impliedThe wordsRead45%−1
Inside the text the surface carriesThe surface it is going toMeasured34%0
Three millimetres of bleed on every edgeThe pressMeasured34%+2
A shape the surface takesThe surface it is going toMeasured34%−3
Nothing set under the surface's smallest sizeTypeMeasured34%+2

Every rule in this table is one that is measured or read, and that is not a coincidence: a rule a vision model raises is a question rather than a finding, and it is not counted here until a person has agreed with it. The same goes for the lines on the checklist. What is left is arithmetic and matching, which is exactly why these are the rules allowed to stop a file on their own.

The period, against the one before it

1 Sep – 13 Sep · against 1 Sep 2025 – 13 Sep 2025

Computed
  • Files checked

    Everything that came through the gate, inside and outside.

    150▼ 0.7% from 151
  • Passed first time

    Nothing measured or read came back broken on the first round.

    69.3%▲ 3.1 pt from 66.2%
  • Sent back

    Each one with the findings and the rule behind them attached.

    46▼ 9.8% from 51
  • Issues per file

    Counted over every file, not only over the ones that broke something.

    0.49▲ 0.04 from 0.45
  • Hours from arriving to a decision

    The gate finishes in under a minute; this is how long the answer waited on a person.

    4.4 h▲ 0.6 h from 3.8 h
  • Rules that caught something

    Out of 58 in the book. This is how much of it earned its place this period.

    24▲ 14.3% from 21
  • Rules that caught nothing

    Not the same as rules doing nothing. Most of these are the ones everybody has already learned.

    34▼ 8.1% from 37

Nothing on this page is an outcome. A quality-control department that reports sales is reporting somebody else's number; this one reports how many rules were broken, by whom, and how long the answer took. Whether the work was any good is a question for the people it was made for, and it is not measured here or anywhere else in this module.

DeskMark-upRulesLive