Brand-constrained content generation

The two constraints are different problems

Voice and fact are usually discussed together and they behave nothing alike. Voice is a matter of taste that a reader notices immediately and forgives quickly. A wrong fact is the opposite: most readers never notice it, and the ones who do are the ones whose opinion you were writing for.

So they are worth solving separately. A voice constraint is a description of how this company writes, derived from what it has already published rather than invented in a workshop. A factual constraint is a set of things the company has said, that a draft may draw on and may not go beyond.

Where a voice constraint comes from

The honest source is the company's own published material. Register, sentence length, whether it uses figures or adjectives to make a point, whether it hedges: all of that is already visible in what has gone out. A house style document written for the purpose describes how a company would like to sound, which is a different thing and usually a more formal one.

The practical consequence is that a voice constraint is only as good as the corpus behind it. A company with two blog posts has not given a generator enough to imitate, and a generator that imitates two posts confidently is bluffing.

Grounding, and what a system should do when it cannot ground

Grounding means tracing each factual claim in a draft back to something in the company's own material. A figure, a date, a certification, a customer count: each either resolves to a source or it does not.

The interesting case is the one that does not. A generator that quietly drops the claim leaves a draft that reads oddly. One that keeps it has invented a fact. The third option, and the right one, is to make a smaller claim that the material does support. "Variance under five microns, measured under the method we publish" is weaker than a number nobody can check and it is worth considerably more.

This is also where the cost of getting it wrong is asymmetric. For a firm whose credibility is technical, an invented statistic in public is not an embarrassment. It is a liability, and it is one that outlives the campaign it appeared in.

What a constraint cannot decide

A brand constraint is not an editor. It can hold a draft to a register and to a set of supported claims; it cannot tell you whether the piece is worth publishing, whether the argument is the right one this quarter, or whether the timing is wrong. Those are judgements, and a system that pretended otherwise would be making the same category error as one that invents a statistic.

It also cannot make a company interesting. Constraining a generator to a voice with nothing to say produces on-brand material nobody reads.

How encra does it

encra keeps a brand profile per workspace: voice, vocabulary, argument style, palette, typography and imagery direction. The writing fields ground every generated word, and the visual fields art-direct every generated image, so a picture and the words it illustrates come from the same description rather than from two unrelated ones.

Claims are checked against the material the workspace has given it. Where a claim has no support, the draft makes a smaller one. Everyday text generation runs on Amazon Bedrock in Stockholm, inside the EU. Image generation is currently processed in the United States, and there is no EU image path yet.