You argued that capability and judgment are separate design problems, and that the second only counts if it can actually stop the first. I build software that way. One of these apps runs today. The other two are a short build away on the same engine.
From your essay: The framework only earns its name if the constraints are genuinely binding.
These apps run on ground, an engine I built. In ground an application is a model first: its states, the moves between them, the rules on each move, and what it needs from the outside world. Code that ignores the model doesn't run. That makes your checkpoints parts you can inspect.
| Your essay asks for | What ground does |
|---|---|
| A constitution that can’t be silently rewritten | Rules are truths, declared by a person. Each change is logged with who made it and why. The machine can hold beliefs of its own, but they are kept separately and can never overrule a truth.ground truth · truth-log |
| A governor that can actually block an action | Every record sits in a state, and the model lists the moves allowed from each state. A move that isn’t listed is refused. The check runs before anything reaches the outside world, so a refused email is never sent.refusal: not_in_state · exit 1 |
| Memory of consequences | Every event goes into an append-only log. Nothing is overwritten. Later decisions can read what happened last time.ground flow |
| Humans govern the permission architecture, not every output | People approve the model and its rules. Individual outputs go to a person only where the model says so. Leads, for example, holds the first ten emails for review and then stops asking.guard: StillInReview |
| Checkpoints that can be inspected independently of the model | Every path through the app is written down as a journey and replayed as a test. Leads has 148 of them. The proof doesn’t depend on how smart the AI inside is.ground verify |
| Uncertainty surfaced honestly | An unclear reply goes to a person instead of being guessed at. When a website doesn’t state a fact, the answer is left empty. When the model doesn’t say what should happen, the app stops and says so.needs_human · model_hole |
Leads researches a company, drafts a first email, follows up, reads the reply and books a meeting. Every step is a move in a model, and the model decides what is allowed.
The refusal format matches the real app: not_in_state names the operation, the states it needs, the state the lead is in, and the moves available from there. The real model has more states and guards than this replay.
Early on, approve_lead sent an email for a lead that was still a draft, and only then refused the move. The rule existed but ran too late. The fix moved the check ahead of every call to the outside world. Now the app refuses before it acts, and a journey proves it on every build. Your essay calls this designing oversight in rather than bolting it on.
Any agent can propose an action. The gate puts it through five checkpoints you define, and the agent can’t skip them. Your essay asks for this: institutional checkpoints external to any one model.
Illustrative run with example actions. In the built app, the constitution and limits are ground truths and guards, past incidents come from the event log, the consequence rating is an ai-text call, and a hold emails you to approve.
You wrote that systems that forget can’t learn from harm; they can only repeat it faster.
The journal records a decision, your reasoning and how confident you are. At 30 and 90 days it emails you to ask how the decision turned out. Over time it shows how well your confidence matches what actually happens.
You warned about governance theater. Each of these apps fails closed. You can make it refuse in front of you, and the refusal comes from the model, not from the AI deciding to behave.
Mail, calendar, web reading, AI text, schedules, counts and lists were built once for leads and are shared by every app. The Gate and the Journal need no new parts, only a model and its rules.
The hard question about WiseEarth is whether it can work in practice. Three running apps, one for governing, one for remembering and one for acting under limits, answer that.
You ended on capability and judgment being separable. Here they are separate: the AI writes, rates and reads, and the model decides what is allowed. Swap the AI for a smarter one and the limits don’t move.
I’ll run leads live, then we’ll write your constitution into the Action Gate together and watch it refuse things. If it holds up, we’ll know what WiseEarth looks like as working software.
Delaney Burke
Zero2One · builder of ground