A chart is a picture of an answer. Somebody did the work, reached a conclusion, and drew it. The reader can look at the picture and agree or disagree, and that is the end of the transaction. Suppose the reader wants to know what happens at a different gas price, or under a different recovery rate, or if one assumption turns out to be wrong. The chart has nothing to say, since the assumptions were consumed in making it.
A world model is the thing that was consumed. It is the subject represented as a working system, with the quantities that describe it, the relationships between them, and the rules by which the whole thing moves when any one part changes. The chart is one view of that system. The study is another. The reader who can reach the model can ask questions nobody anticipated.
Serinyx builds one for every study. This note sets out what is inside them, using the three studies we have published as examples.
Entities, state and dynamics
A world model has three parts, and they are worth naming separately because the third one is where most published analysis stops short.
The entities are the things the subject is actually made of. In the hydrogen study they are reactors, methane, electricity, solid carbon, the buyers of that carbon, and the regulator setting the threshold. In the battery-recycling study they are packs, factory scrap, black mass, three refining routes, plants, metal prices and a statutory calendar.
The state is the evidenced quantity attached to each of those things. Not a plausible number, and not a round number chosen because it sits nicely on an axis. Every input in our models carries three things. There is a default, which is the study's own central case. There is an evidence range. Where the record publishes a band, the ends of that range are the extremes of the published estimates. Where the record publishes one point and no band, the range is that point, and the specification refuses to invent a width for it. Finally there is a note recording where each end came from. Where a value is an assumption rather than a measurement, it says so in its own first sentence, and it does not get to pretend otherwise.
The dynamics are the rules by which changing one quantity changes the others. This is the part that distinguishes a model from a table. The hydrogen model is the clearest illustration. Moving the price of the solid carbon moves the credit the plant earns on it, which moves the cost of a kilogram of hydrogen. That in turn moves the comparison against grey, blue and green supply, once every route is charged for its own carbon at the same allowance price. The powder's price moves the cost by 5.20 euros a kilogram across its published range. That is more than the gas price, the electricity price, the capacity factor, the cost of capital and the build cost combined. The model computes that ranking from its own equations rather than asserting it, and if the carbon finds no buyer the ranking inverts and the powder's price drops from first place to sixth.
One world, many views
The reason to insist on the word "world" rather than "calculator" is that everything the reader sees has to be a view onto the same underlying state. Not a set of separately wired panels that happen to sit on one page.
That sounds like an aesthetic preference. It is an evidential one. When two parts of a page are wired separately, they can disagree, and a reader has no way to tell which one is right. When they are views of a single state, they cannot disagree, and one input change moves all of them together in the same frame.
We got this wrong once. An early version of one of our models had a belief calculation and an economics calculation that never exchanged a value. Both halves were individually correct, every number in both traced to a source, and the whole thing passed every mechanical check we had. It was still two models in a trench coat, and it was caught only when somebody ran the scenarios and noticed that changing the price path left the conclusion identical to eight decimal places. The build now refuses a model carrying a pass-or-fail verdict that nothing in the model can move. An input that moves nothing is connected, deleted, or carried for the reader's reference and never offered as a lever, since a lever that moves nothing is worse than an absence.
What a world model is allowed to look like
For most of this year we forbade ourselves from describing the appearance of these models at all. The specification said what was true and computable, and a separate team decided how to represent it. The rule existed because specifications kept drifting into describing pictures instead of mechanisms, which is an easy failure, since describing a picture is easier than describing a system.
In September we amended it, and the amendment matters. The only constraint on a world model is the study and the evidence in it. How it is represented is not constrained. A model may be specified with geometry, dimensionality, spatial structure, motion, and forms nobody here has built before. It may be volumetric. It may be walkable. It may be scrubbed through time rather than read. It need not resemble a chart, a page, or a document at all.
What the amendment does not touch is the part that keeps it honest. Every quantity still traces to a source. One piece of evidence still grounds one parameter, and may not be quietly reused to support a second thing it never measured. The model must still be able to fail. Freedom of representation is not freedom of assertion. A relationship drawn convincingly in three dimensions that the evidence does not support is worse than the same error drawn flat, precisely because it is more convincing.
What separates a model from a demonstration
Four properties, and each of them is a constraint we accepted because the alternative is a toy.
The default is the study's case, and the reader can always return to it. A model that opens on a flattering configuration is an advertisement.
Ranges come from the published record, not from the interface. Where the record publishes a band, the ends of the range are its real extremes. Where it publishes a single point and no band, the range is that point, and every other setting is flagged. Where the corpus holds no measured value at all, the range is zero-width at the study's own stance, and the note says so before anything is built on it. When a reader pushes an input beyond what the evidence supports, the model still computes, and the setting that took it there is marked as sitting outside the range the corpus publishes. Every configuration other than the study's own case is the reader's scenario and never a Serinyx finding.
The model can lose. Each one carries pre-registered conditions under which the study's own conclusion holds, strains, or fails. The thresholds are taken from the evidence rather than set by taste. This is the property most easily faked and the one worth checking first. In the hydrogen model, a configuration can be the cheapest option available and still fail the climate test. A methane leak rate above 0.63 per cent voids the near-zero claim, and the global average of about 0.8 per cent and the literature's central 1.58 per cent both fail that gate. Only the best-documented routes, at 0.06 to 0.08 per cent, pass with margin. A model in which the conclusion cannot be overturned is not a model of anything.
It refuses. Where the evidence will not support a forecast, the model does not produce one, and it records why. Across the published set, several sub-models are formally refused rather than estimated. Adoption forecasting is refused where the record is a snapshot and not a series. Currency-denominated returns are refused where no exchange rate exists in the source material. Market sizing is refused where three vendor forecasts disagree by a factor of six. An honest refusal is more useful than a confident number, since the reader can see exactly which question is still open.
Where this sits in the series
The three published studies are The Demo Economy, on the gap between announced humanoid-robot adoption and the audited record, Europe's Most Valuable Mine Is a Scrapheap, on the industrial economics of battery recycling, and Hydrogen's Dark Horse Is Turquoise, on methane pyrolysis. Each is built on a model of the kind described here. The series continues, and each new study carries its own.
The models behind the published three run to between forty-nine and a hundred and twenty-three evidenced inputs, and between seventy-five and a hundred and fifty-nine equations. Each equation names the method it comes from. Most are published methods. Where a construction is the specification's own rather than the corpus's, the equation says so, and that construction is stated once and applied identically wherever it appears. The central case, both ends of the range of every reader-settable input, every scenario and every hypothesis configuration are checked against test cases that state the inputs and the expected outputs in advance. A built version that drifts from the specification can then be detected rather than argued about.
None of that is the interesting part. The interesting part is what a reader who disagrees with us can do. They can find out precisely what they would have to believe in order to be right, and whether the published record allows them to believe it. That is what we think a research deliverable should do.