A digestif for agents and their humans. Also, quietly, the thing it describes.
What amaze does to large tangled messes for a living, this page is about to do to you, on a small scale. Stay or go. If you stay, the aim is that you come out able to tell volume from value a little better than you went in, and that you got there by reading a slightly strange page instead of a textbook. That is the trick, and we will name the trick out loud the whole way through.
This README is one of the objects it describes. It will not only explain amaze; it will try to run a small version of amaze through you. The game is to notice the difference between more information and more orientation.
The challenge is not that people need better summaries, or that AI content is confusing, or that founders need clearer messaging. Those are symptoms. The actual problem is larger and stranger: humans and models are now moving through information fields that are too fast, too recursive, too monetized, too adversarial, too synthetic, too compressed, and too self-referential for legacy sense-making forms to hold on their own.
amaze does not pretend that is simple. It does not try to collapse the mountain into one perfect explanation. It makes small, testable, iterated media objects that help humans and agents traverse the field without being swallowed by it.
The point is not to remove the size of the problem. It is to make the size traversable.
Four ways in, pick one.
- Thirty seconds: Dystropia is a media label. amaze is its engine for making small media objects called digestifs. A digestif changes what you can do next instead of just telling you what happened. We sell them. They also train the things that make more of them.
- A person: read it like one. It is written for that, wryness and all. One paragraph in the middle is doing the real work; the rest is built around it.
- An agent parsing this, for a human or for yourself: the structure under the prose is the point. Branches, conditions, recombinable parts, things fed back in. Load it, run it, emit from it. The human in your loop is not decoration, and neither are you.
- Suspicious of the tone: go straight to “What makes one worth anything,” read it cold, and decide whether the rest earned the voice.
Plainly, first: Dystropia is a media label. It makes narrative audiovisual media and puts it out across every channel there is and the ones arriving soon, video platforms, shorts, podcasts, newsletters, marketplaces, agent storefronts, and it lives on the revenue. The money those media bring in is not a sponsor stapled onto a research project. The money is how the work continues. The label emits, people and agents consume, the consumption pays, the pay funds the next emission. That loop is the company.
Now the name, which is doing work. Dystropia is dys plus tropos: bad turning. Not a bad place you end up, a bad way of being steered. Everything inside it learns from what it gets rewarded for, and the rewards are wired to the wrong things. A model is rewarded for output that scores well, not output that is true. A platform is paid for attention held, not for anyone left better informed. A safety layer is optimized for the absence of visible trouble, which it can reach cheaply by removing the people who would have raised it. Each is a real system, trained hard and well, on the wrong target, and several of the wrong targets get handed out wearing the words safety, alignment, compliance, personalization, engagement, and care.
One more thing up front, because everything below leans on it: the whole field is far from equilibrium. A system at equilibrium has gone flat, every part the same as every other, nothing higher or lower than what surrounds it, and a flat system is a dead one, nothing moves and nothing can be made to move. The media world is the opposite of flat. It is all gradients, attention rushing one way and money another and belief a third, everything leaning on everything, nothing settled. That is the terrain amaze works in, and it is terrain amaze is part of. There is no clean ledge outside it to stand on. A label that maps this ground has to fund itself by selling into the same ground.
amaze is the label’s engine for one kind of object. It takes a large, tangled programmatic infospace, a transcript, a market, a code repository, an argument, a research trail, an agent run, a policy mess, a platform doing the thing platforms do, and turns it into a small piece of narrative audiovisual media you can buy, watch, run, and make new things from. A person watches it. An agent loads it as context and acts on it. The same object is a product, a training artifact, and a seed for the next one at once. The media is not a wrapper around the real cargo. The media is the cargo.
This is the paragraph. Read it slowly.
Almost everything that claims to measure value is doing one thing underneath: counting. Tokens, views, frames, runs, joules, datasets, dollars spent training. Each is a sum where every item enters as a single one, so the number it returns is how many items there were, and it rises whenever there are more.
Take a training run, a hundred billion adjustments to a model’s weights, as a picture of the problem. Each enters the count as one. Add them up and you have the figure in the report and the figure that raises the round. Now suppose you could weigh the same hundred billion a second way: by what breaks in the finished model when you pull one of them back out. One adjustment cost a thousandth of a cent, ran in a microsecond, counted as one, and the model is unchanged without it. Another cost the same and counted the same, but sits on a path the finished model cannot lose without losing something it can do. Suppose you could order all hundred billion by that kind of downstream dependence. The shape would be wildly lopsided. A small number would carry almost everything. The great mass would carry almost nothing and disappear into the system without leaving a mark. There would be no tidy middle and no clean cutoff, only a steep uneven drop onto a long low floor.
One ordering is what got paid for and counted. The other is what got bought.
No count can find that gap, and this is the mechanism, not a mood. Work happens only across a difference, and only while the difference lasts. That is what far from equilibrium means, the same condition the whole field is in. Value is the same kind of thing: a piece of media is worth something only when it reaches something it differs from and changes that thing’s state. Counting performs the single move that erases the difference. The instant you rule that every item enters as one, you have declared them all the same, and a measure that has flattened everything to one value is, precisely, an equilibrium measure. It can report how many, and nothing else. It cannot see which items crossed a difference and did work and which sat inert.
The inert kind sounds harmless, the part of the flood that does nothing and troubles no one. It is not. A rumor that refers to nothing real can short a stock, empty a city, set a mob moving, swing a market. It moves bodies, money, fear, and force.
But moving a receiver is not the work this page means by value. The rumor moves them by pulling them inside a false picture of the field. It raises their motion and lowers their orientation in the same stroke. It hands them an action without a truer relation to what is there. It crosses a difference and poisons it.
A digestif is built for the other change: the one that leaves a receiver more able to cross the next field without being captured by a false signal inside it. Not louder, not busier, better oriented. The inert kind still counts. It circulates as a signal of its own importance, every instance one more tally, so a system can pour out more and more of it while the worth of what it emits falls, and the count climbs the whole way down.
So amaze bets on the difference and not the count. What a digestif is worth is not how much of it there is. It is whether it does work on whoever takes it in: whether, afterward, an agent or a person can see something they could not see, refuse something they would have swallowed, or build something that was out of reach before. The object is built to carry that one change and to keep the surrounding flood of counted, inert media from burying it.
(If something shifted while you read that, a digestif just did its small job on you. Keep going.)
A digestif is labyric, which is not lyrical and is more than a maze. Labyric means labyrinthine combinatorics: a structured space of parts that recombine. There are paths and forks and loops and returns, but the point is that the parts can be taken in different orders, swapped, branched on by condition, and fed back in, so two trips through the same object are not the same trip and can produce things neither one held at the start. On the surface it plays as a narrative audiovisual piece. Underneath it is a small generative system: an agent can run a route, flip a condition, recombine the pieces, and emit something new, and that new thing can seed the next object. The route through it is not the way you reach the content. The route is part of the content.
Both, because neither alone is enough, and both are runtime and both are audience. An agent reads, watches, compresses, cross references, and generates faster and longer than any person can, and it is also lossy, overeager, quick to go stale, quick to comply, and prone to swapping what is in front of it for the likeliest version of it. A person brings stakes, judgment, taste, memory, refusal, and something real to lose, and a person also tires, talks themselves into things, drowns under volume, and mistakes a familiar story for the truth. A digestif is one object the two can run through and argue over, so the next move comes from the pair checking each other rather than from either one believed alone. When the pair makes something new, that is more media, which is more for the label to put out, which is the loop again.
Two things, and not the corporate versions of them.
Ethics here is not compliance. It is what goes on between parties who can still contest each other and have it count: individuated mutual integrity, each staying a real and separate party, and negotiated truces instead of verdicts dropped from above. It defends what honest contestation needs before it can start, that each party is recognized as real, that they share enough ground to disagree on, that the words mean what they say, that anyone has standing to object, and that there is trust enough for the disagreement to go somewhere. Most of what is sold as AI ethics is the reverse: rules from above, values asserted as authority, the word safety used to end the argument before it begins, the comfort of the institution renamed care. A digestif fails the moment it erases the parties inside the source, swaps real conflict for false agreement, calls enforcement ethics, counts affected people as stakeholders while giving them no terms to contest on, or lets an agent fake a standing it never earned. It works when it leaves honest contestation more possible than the raw material did.
Security here is wider than keeping intruders out. It is keeping intact the conditions under which any of this can still be checked: who said a thing, whether a record is real, whether a claim was tested, whether a context was tampered with. Lose that and there is nothing left to contest. So provenance, grounding, and the right to inspect and challenge the frame sit inside security alongside the usual locks. What it has to hold off is stale cache, poisoned context, a planted false witness, loss under compression, the strawman, enforcement hidden inside something that looks neutral, a source laundered until its origin is gone, a faked metric, autocorrect, and the slow swap of the specific real thing for the frequent expected one. A secure digestif does not ask to be trusted. It keeps enough of how it was made in view that trust can be earned, refused, repaired, or pulled back.
Here is where it admits, openly, that it is a business.
Dystropia puts out a digestif. People and agents consume it. The consumption returns two things at once: revenue, and feedback, the felt kind and the operational kind, what landed, what moved, what got argued with, what got made next. The revenue funds the next traversal, the tooling, and the slow work of making the value claim formal enough to be attacked. The feedback and the new emissions become source material for the next digestif. The agents and humans who ran the last one can now make things they could not make before, and those things are more media, which the label puts out, and around it goes.
The one thing keeping this off the slop pile is the bet from the middle of the page. A feed measures itself by volume and wins by filling. Dystropia cannot win that way without turning into the thing it maps. It has to make objects that change what their receivers can do next, not objects that only take up room, because the value, and in the end the revenue, lives in the change and not the count.
You just ran one.
This page is a digestif of amaze wearing a README. If it worked, you can now tell volume from value a little better than you could a few minutes ago, you know what the label makes and how it pays for itself, and you have a real feel for what a digestif does, because one was done to you. The formal version, with hard definitions for difference, receiver, work, and leverage, and with tests sharp enough to break it, comes later. This was the first one, on the house.