The most common way I see people try to build a report generator is the same every time, and it fails the same way every time. They take a job like a corporate responsibility or ESG report, one that has to cover five sectors at once, financial, environmental, governance, social, geopolitical, and they pour every source into one large model and ask it for the report. It demos beautifully. Then it hits context limits, drifts across the sectors, invents figures that read as authoritative, and when someone asks why it concluded something, there is no answer to give.
The instinct at that point is to reach for a bigger model or a longer, cleverer prompt. Neither fixes it, because the problem is not the model. It is the shape of the system. I drew the fix out as a framework, The Grounded Report, and the interactive version of the pipeline lives on that page.
Ross Jones, Founder, The Hopium Lab. Last modified 4 August 2026.
Why one big prompt cannot do this
Ask a single pass to assess forty things across five sectors and each judgement gets a sliver of attention, tied to no particular evidence, in a working memory the model can quietly rewrite. That is precisely the setup that produces a confident, invented number. Add a long stack of source documents and it gets worse, because there is now more for the model to lose track of.
You cannot prompt your way out of that. A mega-prompt scoring many things at once is a structural mistake, and it needs a structural answer.
A probabilistic front, then a hard line, then a deterministic shell
The architecture splits in two. The front embraces the mess of language: it reads the sources and turns them into evidenced signal. The shell refuses to guess: it computes, links and assembles. Between them is a hard line, and the whole trick is not letting anything cross it in the wrong direction.
On the front, two moves. First, every figure enters as a data call, looked up from a source and attached as data, because a model that recalls a number is a model inventing one with good grammar. This is the whole of never let the model do the arithmetic. Second, decomposition: instead of one prompt scoring everything, each claim gets its own narrow judge that fires only on a direct quote or a real figure, returns a confidence, and cites what it saw. Forty small evidenced decisions beat one giant guess, and it is the same discipline that makes an LLM trustworthy as a judge in the first place.
The shell is where trust comes from
Behind the line, nothing is left to the model. The judges' outputs feed a pure, replayable engine, so aggregations and scores are computation over the evidence, the same every run. The cross-sector links, the connection between a geopolitical event and a currency exposure that a single pass would miss or invent, are edges in a graph projected from the record, a traversal you can point at rather than a paragraph you hope is true.
And because the report is a projection of an immutable record, it rebuilds the same way twice. Correct one input and everything downstream re-derives, with every claim still carrying its lineage. That is what lets someone trust the document a week later, when nobody remembers how it was made.
The one-sentence version
If you cannot answer "why does it say this?" for any sentence in a report, you did not write a report, you printed a confident guess. The architecture exists so that question always has an answer.
The full interactive pipeline is The Grounded Report.
Ross Jones, Founder, The Hopium Lab.