Maynard⚡️

1/5 Many teams are building memory for AI agents. We studied the question and took the opposite path: our gates are designed to forget. Think of your nose. Spend a while with an odor, in a restaurant, at home, even your own perfume, and you barely notice it any more. Scientists call it olfactory adaptation (also olfactory fatigue, or nose blindness). It is normal, not a flaw. Sensory adaptation lets an organism respond mainly to changes. It is specific to each odor, and sensitivity comes back with time once the odor is gone. The nose stops reporting what stays the same, so it can respond to what is new and out of the ordinary. One study goes further. People were exposed to the same odor. Some were told it was a natural, healthy extract; others, that it might be hazardous. The first felt it fade. The second felt it grow stronger. Yet detection thresholds measured before and after showed the usual adaptation. The sensor adapted either way. What people believed changed what they kept noticing. That is how we design a MONARK gate. It does not try to remember everything. Today it keeps a separate calibration for each class of request it can answer, and abstains on the rest. One of our trackers already works a little like the nose: it corrects its threshold after every outcome, and a daily sentinel steps it and publishes a hash-chained public timeline. Next, upcoming and not built yet: a fixed-size memory for each class, a small summary that forgets on three clocks, fast, medium and slow, so that what stays the same fades into it and what changes stands out. And when the gate does not know enough, it says so: it abstains, and says why, on the wire. Today the gate answers commit, defer or abstain, never a probability of being right. The rest of this thread is what comes next: how errors get paid for, how a bet becomes a trace, what can go wrong, and where we start. 2/5 A colony that fills needs has to try things, and every try can be a false discovery. Run twenty independent tests at the usual 5% threshold on ideas that do nothing, and the chance that at least one looks like a win is 64%. Nobody has to cheat: every choice made after looking at the data is a hidden try. An agent that tests ten recalibrations and keeps the best has not found a remedy. It has found the largest of ten noises, unless each try was paid for. So at MONARK, error is treated as a currency. What already runs in the gate: a running error budget, carried by the caller. When it falls below its floor, the gate abstains, with the reason budget_exhausted. It is a monitoring statistic, not a guarantee, and our design notes say so. What we are designing next (upcoming, not built yet): - a price for every recalibration: each new attempt must use half the error allowance of the attempt before, and a row that does not halve it is refused. This check exists in our code and tests, on synthetic rows; it is not on the served path yet; - a separate wealth of error for discoveries. An agent stakes part of it before testing a new sensor, a recalibration or a merge of classes; - the hypothesis, the stake, the smallest effect that counts and the test data go into the attested log before any result exists, so the data that suggested an idea never judges it; - a confirmed discovery earns wealth back: the abstain zone shrinks by at least that effect while coverage holds, on fresh data. That gain is the agent's credit. No votes. A failure loses its stake and leaves a stop mark, so nobody repeats it; - old wealth fades, so yesterday's lucky streak cannot carry today's tests. If the wealth runs dry, a top-up is a dated human decision, never automatic. The nose spends its attention on what changes. A colony of agents should spend its error the same way: only where a need is real. 3/5 An ordinary trace says: someone passed here. A bet says: someone risked something here, and the world proved them right. Take a hypothesis, say "this calibration still holds", and bet against it, a little at a time. If the hypothesis is true, the game is fair, and a fair game does not make you rich: the chance that your capital ever reaches 20 times your stake is at most 1 in 20, however long you watch and whenever you stop. We checked it on a simulated fair coin: over 20000 games, the capital reached 20 in 4.21% of them. On a coin that lands heads 6 times in 10, every game got there, after 152 tosses on average. Evidence takes time, and the less the world tells you, the longer it takes. That makes a bet the right trace for agents that never talk: - one number sums up the evidence on one question; - a newcomer picks up the bet where the last one left it, on new data, and it is still a valid bet; - agents betting in parallel combine by averaging, with no coordination; - copying someone's trace adds nothing. Only a stake paid on fresh data moves it. And it forgets for the right reason. A bet that restarts every day lets old evidence go, not because it is old, but because the world stopped proving it right. The nose recovers once an odor is gone; old evidence should fade once the world has moved on. Today our code holds one detector of this kind, pre-registered, with its false-alarm budget fixed in advance. It is an instrument and decides nothing. In our daily public timeline, a drift flag only opens a review; it never switches anything by itself. What we are designing next (upcoming, not built yet): - a published, attested capital for each class and each question, its hypothesis and stake written before the data; - daily restarts for anything that can change, with the false-alarm rate declared in advance; - one currency across this thread: a capital that grows against "the calibration holds" justifies an abstain; one that grows for "this fix works" is the discovery that earns error wealth back. The gate keeps its language: commit, defer or abstain. A capital is evidence against one precise hypothesis, never a probability of being right. 4/5 Stigmergy can fail, and nature shows how. Army ants that lose their trail can start following each other in a circle that keeps turning: the ant mill, or death spiral. They can die of exhaustion. The trail is followed perfectly. It just no longer touches the world. People do it too. In an information cascade, it becomes rational to copy those who went before and ignore what you know yourself. The group then rests on the first few, and the cascade is fragile. And there is Goodhart's law, in its common phrasing: when a measure becomes a target, it ceases to be a good measure. Even the nose pays a price. The adaptation that frees it for the new can also dull its awareness of a real chemical hazard. For agents that coordinate through traces, each failure needs its counter: - the spiral: only the world credits a trace; another agent's agreement adds nothing; - the cascade: deposit your signal, not your act, and read every outside claim at its source before building on it; - Goodhart: measure the world, not the counter, on fresh data; - poisoned traces: one observation, one vote, attested; - monoculture: measured diversity of models and sensors; - lock-in: no trail may dominate forever, and some exploration is always kept; - self-fulfilling predictions: watch what the gate does to the world it measures; - a pillaged commons: rules for the shared error wealth, set by those it binds. Today the gate abstains instead of guessing, and we hold ourselves to one rule: every outside claim is checked four ways, at its source, before anyone uses it. Every other counter on that list is upcoming, not built yet. None of these counters removes the risk. They make it visible, and they make it cost.