Every research document has a hidden expiry date, and nobody writes it on the cover. It is the date on which one of the numbers inside it stops being true. A commissioned engagement is delivered, read, discussed, and filed. Some months later a metal price halves, a regulation takes effect, or a plant is commissioned. The document is now wrong in a way that nobody has the time or the source material to correct. The reasoning was sound and the conclusion has quietly expired.
This is not a failure of diligence. It is a property of the format. A document freezes the state of the world at the moment it was typeset, and it does not carry the machinery that produced its numbers, therefore nobody can recompute it. The analyst who could have done so has moved on, and the working files are not in a shape anyone else can pick up.
Serinyx builds each study on a world model, and the model does not have that problem. This note sets out why, using the three studies we have published.
The dates are already in the model
The most direct version of the argument is that a well-built model knows what is coming, because the corpus told it.
Europe's Most Valuable Mine Is a Scrapheap is a study about a regulatory calendar as much as about a process. The battery passport arrives in February 2027. Recovery targets step in December 2027 and again in December 2031. From August 2031 recycled content becomes mandatory, at 16 per cent cobalt, 6 per cent lithium and 6 per cent nickel, rising to 26 per cent, 12 per cent and 15 per cent in 2036. Feedstock arrives on dates fixed years ago by vehicles already sold: 230,000 to 420,000 tonnes in 2030, rising to 1,500,000 to 2,100,000 tonnes by 2040.
Those dates are not decoration in the model. Thirty-three dated events sit on its line, twenty-one of them statutory obligations with a date in the regulation. The recovery step is carried as a driver rather than as prose. Both ends of that driver are the statute's own dated steps, 50 per cent from the end of 2027 and 80 per cent from the end of 2031. A reader in 2027 does not need new research to know what the December step does to the viability of each refining route. They need the model, set to the step in force, and the route gates recompute.
The hydrogen study works the same way. Its projections run to 2035, which is the horizon the source material itself supports. Dated events sit along that line, among them allowance prices, patent positions, plant commissionings, and the point at which the carbon co-product market runs into its own absorption limit. The model reaches that limit at a computed date rather than an asserted one, out of three separately published figures.
Re-running is cheap. Re-researching is not
The second argument is economic, and it is the reason a model is worth commissioning even when a document would have answered the immediate question.
When a price moves, a document has to be rewritten by somebody who understands the whole argument. A model has to have one number changed. The cost of keeping a document current collapses from a research engagement to an afternoon. It collapses because the assumptions were written down in a form a machine can evaluate rather than buried in prose.
Hydrogen's Dark Horse Is Turquoise is built to be re-run in exactly that way. Its cost figures are 5.24 euros a kilogram if the carbon earns nothing, 2.44 euros with the carbon sold at 700 euros a tonne, and 5.39 euros if you pay to bury it. The 5.24 is the adjudicated base case the record publishes, and the model holds it fixed rather than rebuilding a cost stack the corpus never printed. What the reader moves is everything around it. On the cost side that is the gas price, the electricity price, the electricity the plant draws, the build cost, the cost of capital and availability. On the carbon side it is the price of the powder and the share of it that finds a buyer. Each lever moves the cost along the slope the study's own published sensitivity gives it. When the carbon price changes, the comparison against grey, blue and green hydrogen recomputes. When the share of carbon that finds a buyer changes, the ranking of which lever matters most moves with it, and with no buyer the dominant lever disappears. When the upstream methane leakage changes, the climate test recomputes separately from the cost test, and a configuration that was viable can stop qualifying without becoming any more expensive.
There is a discipline underneath this that makes the recomputation trustworthy. Where a model projects forward, it states the tolerance it holds itself to, and it prints how far its reproduction of the known record actually lands from it. One of our published models reproduces the published 2024 allowance average to within 0.09 per cent, and misses the published 2027 projection in the same series by 9.5 per cent. Another misses the one genuine out-of-sample test its corpus allows by 16.44 per cent, widening to 22.77 per cent at the high end of its driver range. Both misses are printed rather than removed by narrowing the assumption until the test passes. Where the corpus is too thin for a reproduction to be a test at all, the model says so rather than claiming a pass. A model that has been tuned until it agrees with history is not evidence about the future.
The model is a deliverable, not an attachment
The third argument concerns what a client actually keeps.
A commissioned engagement ends with documents, and the documents are read once and filed. It can also end with the working model of the subject, which the client's own team keeps using. That is a different kind of asset. It answers questions the engagement never asked, it survives the departure of everyone who worked on it, and it can be handed to a board with the assumptions visible instead of asserted.
We think this matters most in exactly the situations where research is commissioned in the first place. A diligence process where the question keeps changing as the data room opens. A regulatory position that has to be defended eighteen months after it was taken. A capital decision where the committee wants to see what would have to be true for the recommendation to reverse. In each of those, a document answers one question at one moment, and a model answers the question the room actually asks.
The models behind our published studies state their own failure conditions for this reason. Each carries pre-registered thresholds at which the study's conclusion holds, strains, or fails. A team that inherits the model inherits that too, therefore they can see the boundary of the argument rather than having to reconstruct it.
What a model does not fix
We should be plain about the limits, since a model invites over-confidence in a way a document does not.
A model is only as current as the evidence underneath it, and re-running it with new prices does not refresh the research. When the structure of a market changes rather than its numbers, the model needs new evidence, and it should be told so rather than extrapolated through. Every model we build carries a list of limits, stated by the model about itself. It records what it holds fixed that the real world does not. It records which links it chose that the corpus did not, where two of its figures sit on different bases, and where the evidence was thin. That list is written for the person using the model six months later, and it is the first thing they should read.
There are also questions no model should answer. Across the published set, several forecasts are formally refused. Adoption trends are refused where the record is a snapshot and not a series. Returns are refused where the evidence is in euros, the prices are in dollars, and the source material publishes no exchange rate between them. Market sizes are refused where the available forecasts disagree by a factor of six. Those refusals stay in the model. A reader who wants a number there is being told, accurately, that they would be inventing it.
Where this sits in the series
The three published studies are The Demo Economy, Europe's Most Valuable Mine Is a Scrapheap, and Hydrogen's Dark Horse Is Turquoise, and each is a summary of research very much larger than itself. The series continues, and each study carries its own model.
Each of those studies also states how much of the research behind it the reader is holding. That is about one word in a hundred and thirty for Europe's Most Valuable Mine Is a Scrapheap, about one word in two hundred and seventy for Hydrogen's Dark Horse Is Turquoise, and about one word in thirty-six of the finished chapters, or one in seven thousand of the corpus, for The Demo Economy. The document is the part that fits in twenty minutes. The model is the part that keeps working afterwards.