An agentic product can look finished in five different ways.
It can have a screen a person can walk. It can have state that survives the screen. It can touch Stripe, email, DNS, or another system outside itself. It can call a model and record what the model was allowed to do. It can have a beautiful surface that makes the whole thing feel inevitable.
Those are not one truth. They are layers.
The Homebase work made the distinction unavoidable. A live onboarding ritual could be clicked. D1 tables could store real rows. A payment step could look like billing while Stripe truth had not yet arrived. A creature could answer warmly while model truth was still absent or bounded to a different branch. A screen could feel like the product while the first email-to-brain crossing had not yet learned from a real user.
Nothing in that stack is fake merely because it is incomplete. A clickthrough is real clickable truth. A row is real durable truth. A mock payment step can be real dogfood truth. The error begins when one layer is allowed to impersonate another.
Clickable truth means the claim has a path a person can walk. The proof is a route, a button, a transition, a room, a correction box, a place where the promise becomes touchable. Clickable truth matters because it lets the product be embarrassed by a real user before the graph becomes too satisfied with its own idea.
Durable truth means the claim survives the browser. The proof is stored state: a session, an account, an inbox row, a correction, a workload, a migration, an export, a file. Durable truth matters because memory is where a product stops being theater. A creature that cannot remember what changed is a character. A creature whose next crossing changes because of a correction has begun to be a relation.
External truth means the outside world participated. The proof is a processor record, an email delivered through someone else's infrastructure, a DNS route, a webhook, a deploy record, a customer reply, a refund, a support cost. This is the expensive layer in the era of cheap demos. Screens autocomplete. The world does not.
Model truth means a model actually did bounded work and the product can say what happened. The proof is not a fluent answer. The proof is a tracked call, a fallback, a usage record, a permission boundary, and a decision about what the model may infer or act on. Model truth matters because every friendly sentence tries to sound like memory, judgment, and agency at once.
Design truth means the surface clarifies the lower truths without hiding their limits. Beauty is not decoration here. Beauty changes belief. It makes a claim feel safer, more complete, more inevitable. That is why design is dangerous before truth labels exist and powerful after they do.
The rule is simple:
A product claim is true only at the lowest layer where it can be falsified.
If a user can click it, it has clickable truth. If it survives reload and return, it has durable truth. If money moved or mail arrived, it has external truth. If a model acted inside a named boundary and left a trace, it has model truth. If the surface makes the state contract easier to understand, it has design truth.
The layers do not always arrive in order. Sometimes the design sketch discovers the product. Sometimes the database table exists before the interface knows how to expose it. Sometimes an external integration is real before the copy can explain it. The discipline is not to force a universal sequence. The discipline is to label the layer, then cut at the lowest false one.
This is the next form of agentic alignment for a company.
Alignment is usually discussed as if the main danger is a model with the wrong goal. In daily company work, the more common danger is a claim with the wrong truth label. The agent says it built payments when it built a payment-shaped screen. The product says it remembers when it stored local state. The report says the creature learned when a workload row is waiting for a consumer. The dashboard says users exist when the only user is a bypass path. None of these requires malice. The system simply lets one layer speak for another.
That is why the commission is the control surface. A long-horizon agent can disappear into the middle and return with a pile of artifacts that are all partly true. The human cannot retain control by rereading every file. Control comes back when the commission requires each artifact to answer: which truth layer did you advance, what proves it, and what adoption decision follows?
This also explains why the first crossing matters so much. One real crossing can traverse every layer. A message arrives. The system stores it. A model or rule classifies it inside a boundary. The user corrects it. The next crossing changes. The interface shows the change without pretending more than happened. That single loop is small, but it is truth-dense. It gives the company contact, durability, consequence, learning, and trust in one place.
Most agentic products will fail by accumulating unlabeled partial truths. The demo is real, the state is partial, the external system is simulated, the model is decorative, the design is persuasive, and the operator slowly loses the ability to say what has actually happened.
The repair is not cynicism. It is labeling.
Truth labels let provisional work stay useful. They let a dogfood screen be a dogfood screen, a simulated payment be simulated, a model warmup be a warmup, a beautiful mock be a design claim, and a real crossing become unmistakable when it finally happens.
The product is true at the layers where it can be checked.
The company is aligned when everyone can see those layers.