How Caliber writes and scores in your brand.
Caliber writes each draft from the brand’s own rules: its voice, vocabulary, claims rules and legal lines. Words the brand never uses never reach the draft. It then scores the draft against the same guide, one dimension at a time, with the dimensions, weights and guidance range the brand sets, so the score shows its work. Each finding will show the line, the rule behind it and a suggested rewrite, which a person accepts, revises or dismisses. The score informs; a named person approves.
How does a brand guide become a rubric?
A brand guide becomes a rubric when each trait it describes is written as a dimension that can be scored. Each dimension has a weight, a guidance range on a 0–100 scale and lines that show what a high and a low score sound like. The brand sets all of it; Caliber holds no rubric of its own.
The weights add up to 1, and the overall score will be computed from them in code, not taken from a model’s own total. Two kinds of check will sit side by side:
- Word checks, matched in code. Never-words, watch-words and length limits will be matched exactly, in every language the rule lists, and will return the exact words that broke them. A model never decides a word check.
- Reading checks, judged by a model. Voice traits and writing rules need reading. Each draft will be scored more than once, so the result is a range, not a bare number, and the range will be labelled provisional until the scorer has been measured on the brand’s own emails.
The rubric below is the one on our homepage. Holm Coffee is a fictional demo brand.
| Dimension | Weight | Guidance range | Example draft |
|---|---|---|---|
| Warmth | 0.25 | 30–70 | 49, inside the range |
| Clarity | 0.25 | 30–70 | 44, inside the range |
| Origin | 0.20 | 30–70 | 57, inside the range |
| Restraint | 0.15 | 30–70 | 81, past the edge: one finding |
| Craft | 0.15 | 30–70 | 52, inside the range |
A score outside its guidance range is a finding to read, not a failure on its own. The brand voice guide template has a section for each part of a rubric.
How is a finding built?
A finding is built from three parts: the line as it stands in the draft, the rule in the brand’s guide that the line meets or breaks, and a suggested rewrite. The person reviewing the draft then decides what happens to it.
The line will be quoted exactly, so the reviewer can find it in the email. The rule will be the brand’s own, in the brand’s own words. The suggested rewrite will be a line that can replace the original as it stands, in the brand’s voice and in the language of the market the draft is for, not a note about what to change.
| Part | Holm Coffee example |
|---|---|
| The line | “Our best Guji yet, and truly unforgettable.” |
| The rule behind it | Holm’s guide lists “best” as a watch-word: it is used only with proof. Name the taste or the farm instead. |
| Suggested rewrite | “Expect peach, jasmine and a long, clean finish.” |
| The reviewer’s choice | Accept the rewrite, revise it, or dismiss the finding |
Dismissing will be allowed. A finding is Caliber’s reading of the guide, and the reviewer can disagree with it. After a change, the score will say that it was made before the last edit until the draft is scored again.
What can the score do, and what can’t it?
The score informs a decision; it never makes one. It will show how a draft reads against the brand’s guide and where it falls short, and a named person decides whether the email goes to the ESP.
- A range, with its reason. The score will be a range, with the dimensions and findings behind it. When the range straddles the brand’s pass mark, it will say so, and an approver decides.
- It never approves. A score cannot sign off an email, and a good score is not an approval.
- It never blocks on its own. A low score will not stop a draft from being saved, reviewed or exported. The approver reads it and decides.
- Not checked is never clear. A draft that has not been scored will show no score, never a zero, and a check that could not run will say “Not checked”. It is never shown as passed.
- A named person approves. Each market edition will be approved by one named person, on one version, with the time. Nobody will approve work they edited. If the edition changes after approval, it returns to draft.
What makes an edition production-ready?
Production-ready covers the details that are not about voice. Before an edition goes to the ESP, Caliber will check its links, its UTM tags, its images and alt text, the lines the brief says must appear, and the legal line for its market.
- Links. Every link will be checked to see that it resolves.
- UTM tags. Campaign links will be checked for the brand’s campaign parameter, so their clicks can be attributed.
- Alt text. Every image will be checked for alt text, and a placeholder image will hold the edition.
- Must-say lines. Each line the brief requires will be checked in every edition.
- The legal line per market. Each edition will be checked for the legal line the brand’s counsel approved for that market. Caliber never writes a legal line, and an edition whose line is still waiting for counsel will be held, not sent to the ESP with a guess.
The same checks, written for a person to run by hand, are in the multi-market email QA checklist.
What happens after export?
After export, Caliber will give each market a hand-over sheet: the draft’s link, the audience, the sender, the send window in local time and the UTM tags, with a checklist for the steps done in the ESP.
Caliber Email has integrations with Klaviyo, Braze, Salesforce Marketing Cloud Next, Voyado, Customer.io and Iterable.
How does calibration on past emails work?
Calibration is the step that makes Caliber’s review agree with the team’s own. It configures the brand’s workspace; it does not train or fine-tune a model on the brand’s content.
It starts from the brand guide and 6–12 months of past sends, turned into a versioned brand profile with a voice contract for each market. Then 30–50 of the brand’s past emails will be scored by Caliber and judged by the brand’s team, and the rubric will be adjusted until the two agree. The past emails will become the reference lines and the settings in the brand’s own workspace, and nothing outside it.
Calibration will be repeated every quarter against recent sends. A change to the rubric or the scorer will mean the scorer is measured again, and until it is, its scores will be marked provisional.
Terms we use
Market edition
A market edition is one campaign adapted for one market: its language, prices, products, links, legal line and send time, from the same brief and the same blocks as the other editions. Each edition is reviewed and approved on its own.
Legal line
A legal line is the text a market requires in its email, such as sender identification and company details, in the wording the brand’s counsel approved for that market. Caliber will use the line the brand supplies and never writes one.
Brand rubric
A brand rubric is the brand guide written as dimensions that can be scored, each with a weight, a guidance range and reference lines, together with the never-words, watch-words, length limits and writing rules checked beside them.
Guidance range
A guidance range is the span on a 0–100 scale where the brand wants a dimension to sit. A score outside the range is a finding to read, not a failure on its own.
Finding
A finding is one line of a draft marked against one rule: the line as written, the rule behind it and a suggested rewrite, which a person accepts, revises or dismisses.
Named approval
A named approval is one person’s sign-off on one version of one market edition, recorded with their name and the time. A score never approves, and a change to the edition after approval returns it to draft.
Draft export
Draft export is the step that will put an approved edition into the brand’s ESP as a draft.
Questions
Can software score copy against brand guidelines?
Yes, if the guidelines are written as rules that can be checked. Word rules, such as never-words and length limits, can be matched exactly. Voice traits need reading, so their scores carry a range. Caliber is built to do both against the brand’s own guide, and a named person still decides.
Does Caliber train a model on our emails?
No. Calibration configures your workspace; it doesn’t train or fine-tune a model on your content. Your past emails are used to set up your brand profile and to check that Caliber’s review agrees with your team’s, inside your own workspace.
Can a high score approve an email?
No. A score informs the decision, and it never approves an email on its own. Each market edition will be approved by a named person, on one version. If the edition changes after that, it returns to draft and needs approval again.
What happens when Caliber can’t check something?
It will say so. A draft that has not been scored will show no score rather than a zero, and a check that could not run will show “Not checked”. Not checked is never shown as clear, so a reviewer always knows which parts nobody has checked.
Will findings work in every market’s language?
The suggested rewrite will be written in the language of the market the draft is for, in the brand’s voice. A word rule is checked in the languages it lists, so a never-word written only in English is checked only in English. List the local words in the guide too.