Taste as Data: From Brand Guidelines to Verified Signals

There's a value in our image engine's memory that is only allowed to move in one direction.

It works like this. Every time a human at a brand accepts or rejects a generated shot, the verdict accrues into that brand's taste memory. Keep rejecting the adventurous, brand-world-staged compositions, and the system reads the pattern — in the code's own words, the brand "hasn't earned brand-world latitude yet" — and quietly tightens, leaning future renders toward locked, product-first composition.

Here's the part I want you to see: it cannot loosen. Not "doesn't usually" — cannot. The signal is typed so that no code path can ever emit the looser direction; widening it would take a deliberate type change that trips review. The comment on the module says why: "a bidirectional lean would be fabricated signal." We have a record of what a brand's humans rejected. We have no grounded evidence of which way they'd want the machine to loosen. So the machine is allowed to learn caution from data it has. It is not allowed to invent confidence from data it doesn't.

That one-way valve is aesthetic judgment as a behavior — something with inputs you can inspect, provenance you can audit, and limits enforced by a type system instead of a style guide. Taste, in other words, as data.

Which is the conversation I want to have with every founder who tells me — correctly, proudly — "we have brand guidelines." Because you probably do, and they're probably good, and they are about to be read by an audience they were never written for.

Guidelines are written for readers who already have taste

Think about what a brand guidelines document actually is. It's a persuasion artifact for humans. It says "modern but warm." It says "confident, never arrogant." It shows the logo at three sizes and lists two typefaces and asks everyone downstream — designers, agencies, that freelancer doing your Q4 packaging — to interpret.

And interpretation works, mostly, because the reader brings their own judgment. "Modern but warm" lands differently with a good designer than a bad one, and you hire good ones. The document was never the whole system. The document plus a tasteful human was the system.

Now remove the human. Hand "modern but warm" to a generative model and you get the statistical middle of every brand that ever described itself that way — which is most of them. The adjectives don't fail because they're wrong. They fail because they're shared. A general model resolves shared language to the shared average, fluently, in seconds.

So the honest question isn't whether you have guidelines. It's whether anything about your brand exists in a form a machine can consume without interpreting — because interpretation is exactly the step where you become everyone else.

What an encoding actually looks like

In our system, a brand doesn't enter the pipeline as a PDF. It enters as structured identity: hundreds of verified signals, organized into groups — who the brand is, the story and values underneath it, how it's positioned and against whom, who its people are, how it speaks and which words it will not touch — extracted from what the brand has actually made, not from what it aspires to sound like.

The count is honestly the least interesting part. Volume of signals isn't the achievement; anyone can generate a long questionnaire. What makes an encoding machine-consumable — usable by a renderer without a tasteful human standing between — comes down to three properties, and each one exists in our codebase because of a specific failure.

Property one: every signal knows where it came from

The deepest difference between guidelines and an encoding isn't detail. It's provenance. In our canonical store, every signal carries one of three states: machine-generated, human-verified, or absent. And absent is a first-class state — never fabricated over, never quietly filled; the header comment on the core contract calls an absent signal "a labeled gap." Blank and missing are tracked separately, so no consumer ever has to guess which one it's looking at. When our system doesn't know your founding story, it knows that it doesn't know — which is the one thing a general model can never tell you.

Provenance is also what makes completeness an honest metric instead of a vanity one. There's a comment in our confidence scorer I'd frame: "a brand can be highly complete yet 0% verified." A hundred plausible machine-generated answers about your brand is not knowledge — it's a well-organized guess. The score that matters is the verified fraction: how much of the encoding a human who knows the brand has actually confirmed. The system computes that read on every resolve — "your Brand DNA is 61% verified" is a real sentence our store can produce about a real brand, along with an ordered list of which gaps matter most.

That's what "verified signals" means. Not more adjectives. Adjectives with receipts.

Property two: a hard line between what's extracted and what's chosen

Here's the subtler thing, and I think it's the part most AI-branding pitches get wrong: not everything about taste can be extracted.

The clearest example in our system is staging latitude — how much room the brand's encoded personality may take over a scene when the customer's direction doesn't specify. We learned that latitude existed the hard way: an early test render came back as a canyon diorama when we'd asked for a clean product shot, because the encoded identity took over the scene nobody had assigned — a story that gets its own telling elsewhere. The fix was to give the latitude a name — expressiveness — and exactly three levels:

It's defined once, at the single seam where a customer's plain-English direction gets compiled into a render brief; every surface merely selects a level. And anywhere the system fires without a human present, the level is pinned to faithful.

But the decision I'd defend hardest is the one we refused to automate. It would be easy — tempting — to infer a brand's expressiveness from its identity: rugged brand, crank the latitude. The code forbids it, in so many words: expressiveness is "a CHOSEN intent, not an extractable brand property" — the comment adds that a wrong seed just reinstalls the original takeover bug as a default. Extraction tells you who the brand is. It cannot tell you how much room the brand wants a machine to take. That's a judgment call, and judgment calls belong to humans, recorded as data.

And the learning loop honors the same line — that's the one-way valve this piece opened on. The taste memory may walk a brand toward faithful on the evidence of its rejections; moving the other way is reserved for an explicit human choice, forever.

That's taste as data in one mechanism: the encoding distinguishes facts about you, choices by you, and inferences about you — and refuses to let any one impersonate another.

Property three: the signals are inputs, not documentation

The last property is the one that separates an encoding from a very fancy brief: the signals get consumed, mechanically, by the systems that make things.

Some examples from our pipeline. Every product carries a preservation contract — three tiers of visual truth: what generation must preserve, what it may vary, and a strict inner core where, per the comment, "distortions here fail QA outright." Known distortion risks ride along and are consumed as anti-prompt material — the encoding doesn't just describe the product, it actively arms the renderer against the specific ways models tend to mangle it. When a human types a free-text direction like "shot on 35mm film, warm tones," a parser snaps the style-ish phrases to curated presets — and it's deliberately conservative, refusing to match any word that might really be a subject or a scene, because a mis-snapped style is a corrupted signal. When the system proposes style defaults from the brand's identity, those slots are stamped as seeded-but-not-yet-human-confirmed, and the encoding remembers which version of the identity the proposal came from. Propose, stamp, confirm — provenance again, all the way down to art direction. (Color gets the same treatment, with enough of its own drama that we've told that story separately.)

And the encoding has one more consumer worth naming: other people's machines. A brand's guidelines in our system can compile to a machine-readable identity document that outside AI agents can read — gated behind a single explicit opt-in, one guard shared by every surface, closed by default. Which points at where this is all going: the audience for your brand identity is no longer just your team and your agencies. It's every model that will be asked about you, render you, or shop you. Your guidelines have new readers. They read structure.

What this costs, honestly

None of this is free, and I'd rather you hear the bill from me.

Encoding is slower than writing adjectives — most of the real work is verification, a human confirming signal after signal, and that labor is exactly what the verified fraction measures. The provenance discipline means the system will sometimes tell you it doesn't know things a cheaper system would happily improvise — a labeled gap is more useful and less flattering than a fluent guess. And the chosen-intent line means the machine will keep bringing certain decisions back to you instead of making them; that's the design working, but it means "set it and forget it" is not on offer, and anyone who sells you taste-at-scale without a human confirmation loop in it is selling you a machine that invents confidence.

Your guidelines, meanwhile, aren't wrong. They're the human half of a system that now needs a machine half. Keep the PDF. It still does its job for the readers it was written for.

Three questions to ask your own guidelines

You don't need our stack to run this test. Open your brand guidelines and ask what a machine — any machine — could do with them:

That last word is the stakes, so I'll end on it. General models are extraordinary at rendering the middle of a category — that's what they're built from. The only defense is specificity with receipts, encoded where machines can consume it. Because if your brand fits in six adjectives, every AI will render you as the average of everyone else's six adjectives.