Build notes · Grounded AI
The edge of the light
Every financial app has your data. Almost none of them know what they don't know about you. That gap has a name in our system — coverage — and it turned out to be the most important idea we've built.
TL;DR
- 01Coverage is Fogo's model of what it knows — and doesn't know — about your money. Most apps track which accounts are connected; coverage tracks what the system can actually see.
- 02It gates on the five flows of a financial life — income, spending, cash, debt, wealth — not on how many accounts you've linked. Coverage is measured in the currency of your life.
- 03Four states — absent → partial → observed → declared — decide what Fogo is allowed to state, and how hedged. Only the top state earns a flat number.
- 04It's the knob on the grounded dial — the moat. Without it, a generative model is just a more articulate liar.
Why this matters
Three reasons this is the load-bearing idea
Coverage sounds like plumbing. It isn't — it's the piece the rest of the product leans on. Here's what it does for the thing we're building, the thing we're chasing, and the thing we're betting on.
For the product
It's the governor, not a feature
Coverage sits between what Fogo knows and what Fogo is allowed to say. It's the reason you'd trust us over a budgeting app that quotes a figure anyway, or a chatbot that's fluent about no one in particular. Take it out and every other feature is just confident noise.
For the north star
The safety rail under "it already knew"
We're chasing one feeling — the moment Fogo says something true you never told it. Coverage is what keeps that surprise good: a single hallucinated figure about a blind spot inverts the magic into distrust in one screen. No coverage, no north star.
For the business
Generative is a commodity; grounded is the moat
Anyone can wire up a fluent model this year. What's scarce — and compounding — is a system that reliably declines to be confidently wrong, and that knows you better every month. Coverage is what turns raw data into earned confidence a competitor can't buy off the shelf.
The distinction
Having your data vs. knowing your situation
Connect an account to most money apps and a switch flips from off to on. The app now has your data. But "has your data" and "understands your situation" are different claims, and the distance between them is exactly where trust is won or lost.
A tool that has your checking account but not your credit card will happily tell you that you spent $2,100 this month — confidently, precisely, and wrong by whatever you charged. It isn't lying. It simply has no model of its own blindness. Coverage is the name we gave to that missing model: how much of your financial life the system can actually see, and at what fidelity.
The ontology
Five directions money moves
Underneath coverage is a claim about what a financial life is. We model it as five flows — the five directions money moves — and the set is closed: every dollar event is one of them.
The move that matters: we gate on the flow, not the account. A flow is a fact about your life; an account is just a data source. One checking account lights up three flows at once, and no single account lights up all five. So coverage gets measured in the currency of your life — not in how many institutions you've linked.
And a flow is never simply seen or unseen. It's absent (no signal), partial (some signal, known-incomplete), observed (looks whole, no evidence of a gap), or declared (observed and you've confirmed it's everything). Only that last state earns the right to state a flat number. The one before it keeps the hedge — "based on what I can see."
The idea
A model of its own knowledge
Most systems track telemetry: which accounts are connected — a plumbing fact. Coverage tracks something stranger and more valuable: what do I actually know about this person's financial life — an epistemic one. It is the system's self-awareness of its own blind spots. That is the whole trick, because you cannot be honest about a thing you don't know you're missing.
Coverage is the difference between an app that has your data and a companion that knows what it does and doesn't know about you.
The analogy
The doctor, the quack, and the search engine
Bedside manner, encoded
Coverage is a good doctor's awareness of which tests they've actually run. Watch how three characters handle the same incomplete chart:
- the quackstates a diagnosis regardless of what's on the chart — the budgeting app that quotes the figure anyway.
- the search boxhands you fluent, generic WebMD — the ungrounded chatbot, confident and about no one in particular.
- the good doctorforegrounds what the bloodwork clearly shows, withholds the diagnosis that needs a scan they don't have, and names the single test that would resolve the most.
That third posture is the entire product. Coverage is the doctor knowing the edge of their own chart — and that self-awareness is the reason you'd trust them over the other two.
The architecture
Where it sits — the hinge between knowing and saying
Coverage isn't a feature bolted on the side. It's a layer in the middle of the stack — the hinge where what the system knows becomes what the system is allowed to say.
What a financial life is made of
The flows, the classification tiers, the canonical labels. The vocabulary.
What I know
How much of that vocabulary I can observe for you right now — absent, partial, observed, declared.
What I say
What I'm therefore allowed to state, and how hedged — what to lead with, what to soften, what to withhold.
That I know you
Coverage and understanding accrued over time — the relationship, not the snapshot.
Everything above coverage is raw capability. Everything below it is expression. Coverage is the governor in between — and an observability layer that recomputed the thing it observes would be worse than none, because the day its copy of the logic drifts, it reports a state the system never actually took.
Why it's the moat
The knob on the trust dial
We think about the product as two dials. One is generative — a system that can say almost anything, richly, in your situation's own terms. The other is grounded — whether what it says is true of you. Coverage is the knob on that second dial.
Without coverage, "generative" is just a more articulate liar.
Generative richness is cheap now; anyone can wire up a fluent model. What's scarce is a fluent system that reliably declines to be confidently wrong. Coverage is what turns raw data into earned confidence — which is what makes "grounded" a real moat instead of a slogan.
It's also the safety rail on the thing we're actually chasing: the moment a new user feels "it already knew." You cannot land that moment if the system is confidently wrong about a blind spot — a single hallucinated figure inverts the magic into distrust in one screen. Coverage is what lets the good surprise happen without the bad one.
Keeping honest
Three ways the idea can mislead you
A concept this load-bearing is also easy to over-trust. Three tensions we hold on purpose:
i. Observability isn't understanding
Seeing all five flows is not the same as knowing what they mean to you — whether that $3,000 is rent you resent or a home you love. There's a layer of meaning above coverage that the flows don't capture. The trap is letting "coverage" quietly redefine the goal as completeness ("connect more accounts!") when the goal is understanding. Coverage is the floor, not the building.
ii. It's private epistemics — don't hand it to the user as homework
Coverage is how the system reasons about itself. Exposed raw, it becomes a deficiency scoreboard — a checklist of everything you haven't done. The discipline is to turn "here are your gaps" into "here's what I can tell you, and the one thing that would unlock the most." Coverage should be felt as confidence, never shown as a chore.
iii. It only means something attached to a real, consenting life
A system can be fully "covered" and describe no one — we watched a couple hundred throwaway test connections carry perfect coverage and mean nothing. Coverage without a real, consenting person behind it is just plumbing that runs to a ghost.
Coverage is the self-knowledge layer. And self-knowledge — knowing the edge of your own light — is the precondition for trust.