Evidence That Survives Scrutiny

A policy decision is judged twice. It is judged when it is taken, on whether it is sensible, and it is judged later, on whether the evidence behind it was sound. The second judgement is the harder one, since it happens in front of an auditor, a committee, a court, or a journalist, and it happens after the people who assembled the evidence have moved on. What survives that second judgement is not the quality of the reasoning. It is whether each figure can still be traced to a source that a stranger can open.

Most research does not survive it well. Numbers arrive without provenance, a widely repeated figure turns out to have one origin, and the gaps in the record are silently filled with plausible sentences. Serinyx built Quorum to produce research where those three failures cannot pass unnoticed. Today we published three reports that demonstrate it, and two of them are on questions European institutions are actively deciding.

This report follows the industrial economics of battery recycling under the recycled-content regime. Recycling feedstock is expected to reach 230,000 to 420,000 tonnes in 2030, rising to 1,500,000 to 2,100,000 tonnes by 2040, on dates already fixed by vehicles sold years ago. From August 2031 recycled content becomes mandatory, which means the law rather than the metal price becomes the buyer.

That is the point at which the report becomes uncomfortable, and usefully so. The metal the law made valuable has become cheap, and a seller of black mass keeps only three to five per cent of the lithium price. The refining route decides whether the lithium survives at all, since smelting loses most of it to slag while hydrometallurgy and direct recycling retain almost all of it. Announced European refining capacity for 2030 stands at roughly double the top of the expected feedstock band, and some existing plants report utilisation of about ten per cent. Nothing in the regulation requires the recycling to happen in Europe.

The report does not tell you what to do about that. It sets out what the record supports, which is a different and more useful thing for a policy team to be handed.

This report models methane pyrolysis, which produces clean hydrogen and solid carbon instead of carbon dioxide. It needs roughly 12 kilowatt hours of electricity per kilogram against 52.4 to 55 for electrolysis, although on total energy it is the hungrier route, since it also consumes methane. What it saves is electricity, and only electricity. That distinction matters a great deal to anyone allocating grid capacity, and it is routinely lost when the route is described in a single sentence.

The economics turn on whether the solid carbon can be sold. A kilogram of hydrogen costs 5.24 euros before any carbon revenue, 2.44 euros with the carbon sold at 700 euros a tonne, and 5.39 euros if that carbon has to be buried. Separately from cost, a methane leak rate above 0.63 per cent invalidates the climate case, and both published figures fail that threshold. A configuration can therefore be the cheapest option available and still not qualify for support. For an agency designing eligibility criteria, that separation between the cost test and the climate test is the whole question.

The Demo Economy is about announced adoption in humanoid robotics. It is less directly a policy document, although its central finding travels. The most-cited deployment metrics in that field appear twelve times across the reviewed record. They collapse to a single origin, which is the vendor's own pilot dataset.

Three properties matter for a document that will be scrutinised after the fact, and each of the three reports carries all of them.

Every figure resolves to a source. Each report carries a numbered reference list in which every link was fetched and verified before publication. Sources that cannot be publicly linked are listed as such and say so. No placeholder or reconstructed links are used. Where a source cannot be published, the list says so rather than substituting a secondary link that can.

Corroboration is counted in origins, not in appearances. A figure repeated by twelve outlets that all took it from one press release is one source. Treating it as twelve is the most common way a policy evidence base becomes quietly circular, and it is the failure The Demo Economy was chosen to demonstrate.

Absence is recorded as a finding. Where the record is silent, the reports say so rather than estimating into the gap. The Demo Economy finds 292 disclosed award projects for 2025, one fully documented acceptance chain, no verified payment remittances, and no verified repeat purchases anywhere. It then declines to over-claim, since publishing those deeper stages is not generally mandated for ordinary goods procurement in China, therefore those zeros are ceilings on what can be known rather than statements about what happened. That sentence is more useful to a regulator than a confident number would have been, since it identifies a disclosure gap that regulation could close.

Each report is also built on a working model of its subject. Every input in those models is bounded by the extremes of published estimates rather than by round numbers chosen for effect, and the conditions that would overturn the conclusion are stated explicitly. An accompanying interactive version lets a reader change the assumptions and watch the conclusion hold, strain, or fail. For a consultation response or an impact assessment, being able to show the committee exactly where a finding breaks is worth more than the finding itself.

Each report states on its closing page how much research it summarises, and we think those figures are the most useful thing in the series.

The battery-recycling report runs to roughly 7,000 words and summarises one subtopic in a research programme of 12 fields and 72 subtopics. That subtree alone holds approximately 913,000 words across 285 documents and datasets, which makes the published report about one word in a hundred and thirty of the research behind it. The hydrogen report draws on approximately 2.08 million words across 1,051 documents and 16 research domains. Those domains run from reactor engineering and the patent landscape through to permitting and funding. The published report is about one word in two hundred and seventy of it. The robotics report summarises a study whose working corpus runs to about 50.1 million words in 29,111 files.

A twenty-minute document is not the same instrument as an evidence base. The value of publishing the ratio is that a reader can see which one they are holding.

We should be plain about the limits, since a public demonstration invites the wrong conclusion.

These reports answer questions we selected rather than questions a ministry set. They stop at the depth a general reader will follow. They cannot draw on client data or anything held under confidentiality, and sources that cannot be publicly linked are listed without a URL and say so. They compress a body of research into a fixed length, which means most of the sub-questions and nearly all of the working analysis stay behind the summary.

None of this is a limit of the pipeline. It is the shape of a document written to be read in twenty minutes by someone outside the field.

A commissioned study starts from your question and your scope, agreed before any collection begins. The evidence base is built for that question at the scale the question needs. Your own material can sit inside it alongside the public record. The review before release is adversarial by design, since somebody is asked to argue against the conclusion using the same evidence. The deliverable follows the decision rather than a template. It might be an impact assessment input, a regulatory analysis, a state-of-the-evidence review, an option appraisal, or a model your team keeps using after the work is delivered.

The three reports are free to read in full, and we would rather be judged on them than on a description of our method.