Your own model cannot validate your own book.

An underwriter's internal model is the right tool for deciding what to write and at what price. It is structurally the wrong tool for one specific job: proving that book to somebody else. A reinsurer asked to take a share, a capital provider funding it, a rating committee assessing it or an auditor signing it cannot validate a book with the model that wrote it. This is not a criticism of anyone's modelling. It is a property of who built it, and it is about to matter in carbon.

The pattern nobody expected in this class

In property catastrophe the modelling layer stayed independent. The science was too expensive for any single carrier, and a shared language turned out to be worth more than a proprietary edge. Carbon has gone the other way, at least so far, because the tooling is cheap and the market is small.

The result is visible in public announcements. New carbon underwriters launch with modelling and monitoring stacks as table stakes. Others pair with external data and rating providers, integrating third party project assessments directly into their underwriting platforms. One carbon insurance vehicle now runs its own satellite and machine learning monitoring inside a broker facility.

Every one of those arrangements supplies the same thing: a view formed before the fact about whether a credit is likely to be what it claims. That view can be built in house or bought in, and increasingly it is both. What none of those arrangements supplies is the other half.

The two questions, and why only one is being answered

Underwriting a carbon position requires answers to two different questions.

What is true now? Is the project real, is the boundary defensible, is the methodology sound, is the forest standing this month. Monitoring, ratings and satellite analytics answer this, and they answer it increasingly well.

What happened before, and how often were we wrong? When credits of this type, this vintage, this geography and this risk profile were written previously, how often did they reverse, get invalidated, lose eligibility or fail to deliver, and by how much. Nothing in the current stack answers this, because the record it would require has never been assembled.

The second question is the one that gets asked when the book leaves the building.

Why an internal book is not a market record

Suppose a carbon underwriter has been writing since 2022 and has three years of claims experience by 2028. That experience is real and valuable, and it is still not a loss table. Three reasons.

It covers only the risks that carrier chose to write, in the years it wrote them, at the prices it set. Selection is baked into every observation.

It cannot say how the risks that were declined performed, which is exactly the counterfactual a reinsurer needs to judge whether the selection was skilful or lucky.

And it cannot be published without exposing pricing. An internal book is commercially confidential by construction, which means it can never become the shared reference that a market prices against.

A book is not a market record. It is a sample chosen by the person being assessed.

Where independence is worth the most

Independence in this market is usually discussed as an integrity virtue, which undersells it. Its commercial value is concentrated at one point in the chain: the moment a book is sold on.

That moment is arriving in carbon now. Capacity is being expanded and syndicated. Strategic investors are taking positions in specialist carbon underwriters. Lenders have advanced non recourse credit secured on carbon, in a structure with a named public precedent. Every one of those transactions has a party on the other side who must form a view about a book they did not write, using something other than the writer's own model.

The catastrophe market solved this with vendor models validated against industry loss data, and with an independently published event loss number that contracts could trigger on. Neither of those was produced by a party to the trade. That was the point.

What the independent layer actually has to do

Three things, in order, and none of them requires displacing anyone's internal model.

Publish the outcome record. Every realised loss, dated, typed and resolved to the document that decided it, free to read and cite. It is the test set the whole class currently lacks.

Run in parallel, not in replacement. The useful proposition to an underwriter is a benchmark: your model and an external read on the same submissions, compared over twenty risks. Nobody has to lose an argument for the comparison to be worth having, and after twenty submissions the comparison itself is the document.

Answer accumulation at the book level. Reversals are physically correlated by fire weather, drought and regional policy. A portfolio of forest projects in one basin is not diversified in any useful sense. That is a question about the whole book rather than any single risk, and it is the question a capital provider asks first.

The uncomfortable sentence

If you write carbon risk today, your analytics are probably better than the market gives you credit for. The problem is not quality. It is that the person who has to take your book cannot check it with your tools, and there is currently nothing else for them to check it with.

That gap is not going to be closed by anyone selling credits, and it cannot be closed by a carrier's own model. It requires a record that belongs to no one in the trade.