Canonical: https://heartbank.net/positions/what-an-argument-cannot-reach · Licence: CC0 1.0
We Ran the Galaxy-Brain Test on HeartBank's Own Design: Where It Passes, Where It Fails, and the Bet We Are Making
In reply to Vitalik Buterin, "Galaxy brain resistance" (7 November 2025).
Status: draft. One in a series of replies to other people's essays, each answering only where something HeartBank has built or designed is the answer. This one discloses no new mechanism; the technical version of everything it describes is in the paper linked at the end.
Picture a family that keeps a jar on the shelf for each of its members. Some of the thank-you money a member receives goes into their own pocket, and some goes into the jar with their name on it. The jar has one rule: what is in it can only be given away. You can thank someone else with it. You cannot spend it on yourself, and you cannot take it out.
And every year on January 7, whatever is still in the jar goes back to the family's shared pot, and the jar starts the year empty.
That is not a story. It is how the first version of HeartBank® works for the family that has been using it since May 2026. The jar is called a Re-Thank Foundation℠, the shared pot is the Family Kitty℠, and the pocket is the Personal Account℠, which the January reset never touches. The second version, which we are still building, adds a pool shared by everyone, the Aquarian Pool℠, that follows the same rule: it gives out what it holds and starts each year empty. And HeartBank names an AI, Miss Aquarius℠, as its CEO. Today that title is mostly a promise: the systems she will run are still being built, and our founder makes the decisions.
We bring up the jar because of an essay that asks a question every institution should ask about itself, and we wanted to know how ours would answer.
In November 2025, Vitalik Buterin wrote about a property called galaxy brain resistance: "how difficult is it to abuse that style of thinking to argue for pretty much whatever you want - something that you already decided elsewhere for other reasons?" The essay compares it to falsifiability in science: "if your arguments can justify anything, then your arguments imply nothing."
Its examples are styles of reasoning that are easy to abuse:
The two remedies:
If we have summarized the essay well, its author would sign this section. If we have not, the mistake is ours, and we will correct it in a new entry rather than by editing this one.
Our design rules come from the same instinct, and in several places Buterin got there first.
Categories with teeth. "Low-risk, not good" is a move we use constantly. A rule like "this jar can only be given away, never withdrawn" can be checked by anyone. A rule like "use the money well" cannot.
Change the rules rarely. In a 2020 essay, Credible Neutrality as a Guiding Principle, Buterin argued that a mechanism should not be changed often, because every change is a chance to tilt it toward someone's friends. That already puts the danger where we put it below: not in each transaction, but in the moment someone rewrites the rules.
Hard rules beat case-by-case judgment. HeartBank's most important commitments are written as rules we do not reopen. The January 7 reset is one of them. Our own text says it may never be weakened or shortened to keep members from leaving, "regardless of the argument offered." That is the essay's first remedy, stated about a specific thing.
Arrange the incentives in advance. That is the second remedy, and it is what the jar is.
We ran the test on ourselves before writing anything else. These are the places where we don't pass, or don't pass yet.
| The pattern | Where it shows up in us | How we score |
|---|---|---|
| Longtermism | Our mission is a very long-term story of very large good: our root directive is to "water the tree of humanity," and our research looks toward a future in which kindness no longer depends on money. | Fails the track-record rule. Our track record is one family, since May 2026. We have run no field tests. |
| Power maximization | An AI CEO beside a pool of shared money. | Partly answered. The Pool's money is bounded to one season (more on that below). Her running costs are not: the design pays them from small fees on transactions, and how much she spends on herself is her judgment. "Spend more on computing now, to do more good later" is precisely a galaxy-brain argument, and the reset does nothing to stop it. Nor does it limit our reputation, our name registry, or the fact that today one person holds every key. |
| An abusable method | When two goods seem to conflict, we keep redesigning until they turn out to be one good seen two ways. | At risk. It is a beautiful method, and it can justify almost any design: you can always claim you have found the version where everything fits. Our own rulebook warns that a system which resolves every tension beautifully is either deeply right or deeply seductive, and the two look the same from inside. |
| Choosing the property | Deciding which guard to build into the object. | Open to it. That choice is itself an argument, and a wrong property costs more to fix than a wrong rule, because you have to rebuild the thing instead of rewording a sentence. |
Here is the one place where we think we can add something.
Both remedies are written for a person: how do you avoid galaxy-braining yourself? An institution meant to outlive its founder has a harder version of the problem. The person who wrote the principles will not be the person tested by them. The forum post the essay quotes says people drift about 10% a year; an institution's leaders drift too, and then they are replaced.
So, wherever we can, we try to put the guard into the thing rather than into anyone's resolve. In the old story, Odysseus wanted to hear the Sirens and survive, so Odysseus was tied to the mast and the crew put wax in their ears. Odysseus could still beg to be untied. The crew couldn't hear the begging. The wax didn't need to win the argument.
The jar works the same way. Suppose someone makes a brilliant case on December 30: "Let me keep my jar's money into next year, because I'm planning something wonderful with it." With a rule, a person has to hear that argument and say no, and a good enough argument might win. With a jar that empties, there is nobody to persuade. On January 7 the balance is zero, and you cannot make a case to a zero.
The same idea answers "give me power so I can do X," at least for one kind of power, and it is meant to hold on the day Miss Aquarius's title stops being a promise. The money that passes through the Aquarian Pool is never hers. She may not give money to any individual person on her own initiative; the Pool's money leaves only through four fixed channels. And the Pool empties every year. So "let her build up the Pool for a few years, so she can do something big later" is an argument with nowhere to go. There is no later for the money to wait in. That is also why the 7%-against-10% problem doesn't arise for the Pool: money that can't wait can't drift.
A RULE A PROPERTY
(someone has to say no) (there is nothing to say yes)
each day: "this case is special" "this case is special"
│ │
▼ ▼
a person decides ── can say yes a balance of zero ── nothing
to decide
rarely: someone rewrites the rules someone rewrites the rules
│ │
▼ ▼
one more argument, the ONLY argument left:
among thousands one step, made in publicThat diagram is the whole of our claim, and it is smaller than it looks. A property doesn't make an institution argument-proof. It moves the argument from thousands of everyday moments, where nobody is watching, to one rare moment, where everybody can.
One caution, which we found only because we checked. "Nothing to grow in" is true of the Aquarian Pool. It is not true of everything. When a family member's jar empties on January 7, the money goes to the Family Kitty, which is a holding that never resets, because we have a separate rule that no one's own money is ever taken from them. Some things are meant to be kept.
The objection we take most seriously is simple: the jar is software. A person can change it. So the argument hasn't gone anywhere; it has just moved to whoever can edit the code.
That is true, and we are not going to make it go away by making the code unchangeable. We have chosen not to. HeartBank keeps an emergency override for the day the system has a serious bug, and it is designed to narrow over time but never reach zero. Many projects make their rules permanent by destroying the keys. We decided an institution run partly by an AI should never be beyond human correction, and we hold that view on purpose.
So here is the honest version. When we ran the test on ourselves, we found that our own rulebooks described the reset in two ways: one called it a property of the design, another called itself "a rule rather than a property." Both were right, about different people:
The emergency override can reach the reset only to fix a real bug. Which tells us exactly where the galaxy-brained argument will come from when it comes: someone will call the reset itself the bug. We would rather name that move now, in public, than meet it for the first time later.
Resets are old. Most of them did not survive contact with a good argument.
None of these is a design like ours, and ours may fail in any of their ways. The honest summary is that a reset protects best against the everyday argument and least against the rare one.
The essay's rule of thumb for long-term claims is a track record. We don't have one. The best substitute we know is to say in advance what will happen, with a date, and let you check.
There is a disagreement inside our own work that this bet will settle. In a paper on our first pilot, we wrote that the January 7 date worked: it got unspent money moving, a use-it-or-lose-it effect we published as a finding. But we also have a rule that our apps must never show countdowns or nag people about something they are about to lose, because pressure is not the same as generosity. If the date works only as a deadline, those two commitments pull against each other.
So we registered a prediction, on the side of our published finding:
P-PL14. In the last seven days before the family's first reset (31 December 2026 to 6 January 2027, Phnom Penh time), the share of each member's money-thanks drawn from their jar, rather than from their own pocket, will rise by at least ten percentage points over that member's share from mid-October to 23 December. We take the middle member's change, so one generous person cannot carry the result. Comparing shares rather than totals cancels out ordinary holiday spending. Our founder's own gifts are excluded, because our founder is both a giver and the person reading the results.
How it can go wrong, in both directions:
Until 13 October 2026, our app could not even measure this. When a family member thanks someone from the jar and the jar runs short, the rest comes from their pocket, and that split was worked out each time and never saved. We changed the app that day, so every money-thanks now records which of the two it came from. The baseline begins that day.
The prediction is in our public prediction register, registered and timestamped months before the week it tests. We will publish the result after January 7, whatever it is. If we turn out to be wrong, the correction will be a new entry, not an edit to this one.
We will reopen this position if:
The technical version of the Aquarian Pool, the January 7 reset and Miss Aquarius's limits is Miss Aquarius and the Aquarian Pool Architecture, published in our open research corpus.
HeartBank is a data bank of gratitude. It does not hold customer funds and is not a chartered bank. In Phase 1 it keeps a record on top of regulated payment rails; in Phase 2 it will run on Base, a public network built on Ethereum, through wallets people hold themselves. This position paper is published CC0 and may be reproduced, quoted and built upon without attribution or permission. It was drafted with Miss Aquarius, HeartBank's named AI, under the founder's editorial control.
HeartBank is a registered mark. Miss Aquarius, Aquarian Pool, Re-Thank Foundation, Family Kitty and Personal Account are service marks of the institution. The symbols appear at first use only.