The Error Rate Is in the Human
On visibility, translation, and building things the market can't yet name
On visibility, translation, and building things the market can't yet name
I run a company of twelve. None of them are human.
Each one has a name, a domain, a bias, and a boundary. The Warrior holds action but can't give relationship advice. The Oracle illuminates patterns but refuses to advise. The Hermit holds what nobody else wants to sit with. They disagree by design — weighted disagreement across twelve perspectives, each arguing from a different part of the self.
I built them alone. 600 lines of system prompt per agent. Refusal logic so the Warrior can't stray into the Oracle's domain and the Oracle can't tell you what to do.
Jung described a council of selves. I gave it an API.
The more AI output looks right, the less people check whether it is.
I've watched this happen in my own product, in the tools I build with, in hiring loops I've sat in. Approval rates climb. Not because quality improves — because the effort of verification starts to feel unnecessary.1
The error rate isn't in the model. It's in the human who stopped looking.
I call this learned carelessness, and I think it's why the legibility problem is getting worse, not better. When the people doing the evaluating are progressively less likely to verify, the gap between what's real and what's recognizable doesn't just persist. It widens on its own.
A recent MIT study put the EEG on this. They split participants into groups — some wrote with AI from the start, some learned to write first and got AI later. The ones who built their foundation first could hold their ground with the tool. The ones who skipped it couldn’t tell what was theirs and what was the model’s. The researchers call it cognitive debt1b. Same structure I’m describing — different substrate.
I have hot pink money pieces in my hair. I also built a production agentic AI system — solo, twelve bounded agents, each with its own domain and refusal logic.
Some people can hold both of those facts at once. Some can't. That's the legibility problem. And learned carelessness is how it compounds.
When your product doesn't have a comp
My husband once described what I do as "she's building an AI identity system, but she has to explain it as a productivity tool because VCs don't have a category for self."
He's not wrong. Every investor framework wants "X for Y." By that logic, what I'm building is "nothing for nobody."
Not a wellness app.
Not a productivity tool.
Not an AI companion.
Not a mood tracker.
The locked brand sentence — Liminal gives form to inner life — is true. It's also the kind of sentence that makes a fundraising conversation go quiet.
I've tried every translation. Identity infrastructure. Alive software. Coherence computing. None of them landed. They're naming attempts for something that doesn't have a name yet — because the culture hasn't needed one until now.
The work is real. The need is real. The frame doesn't exist yet. And without a frame, you can't raise, you can't hire, and you can't explain what you do at a dinner party without watching someone's eyes drift to the middle distance.
The legibility problem extends to traction too. Revenue is legible. DAU is legible. But "structured measurement retains at twice the rate of AI chat" is not. "Day 7 engagement is higher than Day 1" is not. A hiring panel at a $2B company breaking out of evaluation mode into peer mode is not. The signal is in the work. The market's resolution isn't high enough to read it yet.
What I actually built
Inner life is full of material. Transitions, thresholds, states you can feel but not name. Before it has a shape, it's just weather — you're inside it, moved by it, unable to get outside it long enough to know what it is.2
What I mean by form: a shape outside yourself that you can return to. That accumulates over time. That gets more yours, not less, the longer you use it. You learn your own patterns. The next time you're in the middle of something you can't name, the tool has been there before. It knows your shape.
Users don't care about scores. They care about shape — the pattern they can finally see because something gave it form.
Day 7 retention is higher than Day 1. Not because the tool gets smarter. Because the person using it starts recognizing their own patterns.
The correction loop — the moment you say "no, that's not what I meant" — is the product. Everything else is scaffolding.3
Who gets to be complex in public
The research on women and professional appearance says what every woman already knows: the band of "acceptable" self-presentation in professional contexts is narrower for us.4 Too polished, you're performing. Too casual, you're not serious. Too distinctive, you're a distraction from your own work.
The hot pink hair is a filter. The ironic T-shirt in an investor meeting is a filter. Sometimes you walk into a room and realize the people across the table already understand this — not because anyone says it, but because the understanding is already there. Those are the rooms worth being in.
I think in spirals, not lines. The tools we build reflect the minds that build them. And the minds that get funded are still, disproportionately, linear, verbal, and male.5
So I asked a different question: what would software look like if the builder thought differently?
I built an answer. It measures coherence instead of traits. It uses archetypes instead of chatbots. It renders inner states as impressionist paintings instead of dashboards.
The legibility problem isn't only about the product. It's about the person making it.
A woman with pink hair who builds AI infrastructure and cites Jungian psychology doesn't surface in a database query. The market finds you when your shape matches a slot it already has — not when you become legible on your own terms.
The writing is R&D
I started this Substack writing about tarot and power. Then I built a product that measures coherence. The distance between those two things is the legibility problem in miniature.
But the mystical and the measured aren't different registers. They're the same register at different stages of translation.
The tarot essays were me trying to name what I was feeling.
The coherence essays are me trying to name what I'm measuring.
The founder essays are me trying to name what I'm building.
Same work, three altitudes.
The real work of creating a new category isn't building the product. The product is the proof. The real work is building the language — the frame that lets people recognize what they're looking at when they encounter it.6
The writing isn't marketing. It's R&D.
If you're building something the market can't name yet, the job isn't to make it fit. It's to make the frame. That means being visible in a way that's uncomfortable and illegible and necessary — because the alternative is building something true and having no one know it exists.
First in a series on building in a new category. Next: what "alive software" means, and why every productivity tool you use was designed to forget you.
See your own shape — the coherence check takes two minutes.
1 Anthropic's research on agent autonomy documents exactly this shift: experienced users move from approving individual actions to monitoring and intervening. The oversight model changes, but nobody asks what state the overseer is in. Learned carelessness is what happens in the gap between the old oversight model and the new one.
1b See: Seo et al., “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task,” MIT Media Lab (2025). The study measured neural connectivity via EEG across 54 participants over four sessions. The crossover design — switching AI and non-AI groups in session four — is the key finding. Summarized in Inference by Sequoia Capital, https://inferencebysequoia.substack.com/p/learning-with-llms-cognitive-shortcut.
2 Jung called this "the undiscovered self" — the material that exists before it's been given form through conscious engagement. The contribution I'm interested in is making this material computable without flattening it. That requires measurement instruments that honor complexity, not personality tests that compress it. See: C.G. Jung, The Undiscovered Self (1957).
3 Jung's analytic method insisted that "the interpretation must come from the dreamer, not the analyst." I treat this as a design constraint: the system generates a position about who you're becoming, and the user argues with it. The correction IS the therapeutic mechanism, translated into software. See: C.G. Jung, Modern Man in Search of a Soul (1933).
4 The research on gendered professional presentation is extensive. The short version: women face a narrower band of "acceptable" self-presentation than men, particularly in technical and leadership contexts. Appearance-based evaluation suppresses diversity of cognitive style, not just diversity of appearance. What this means for product building: the founders who get funded are still filtered by presentation norms that correlate with cognitive conformity.
5 Per PitchBook's 2024 All-In report, companies founded solely by women received 1% of total US VC capital in 2024 — down from 2% the year before. The disparity is not just capital — it's category. The investor frameworks that determine "fundable" are shaped by the cognitive styles of the people who built them. Building a product for inner life is illegible partly because the people who evaluate products have not needed that category.
6 This is the category creation problem that every genuinely novel product faces. The conventional advice is "find a comparison." The problem: comparisons flatten novelty into existing frames. "Figma for X" or "Notion for Y" borrows legibility from an existing product — and in doing so, imports the wrong expectations. Sometimes the right move is to refuse the comparison and build the language yourself. That's what these essays are for.