The architecture
One source analysis. Three reading surfaces.
The product rests on one rule: a report is never written straight from raw sources. Each surface is drawn from an analysis that already exists, so a company is read in depth once rather than three times, and every surface agrees with the others.
The canonical artefact
A structured due-diligence file, every statement a claim.
The source analysis is produced once per company and holds everything the surfaces draw on. It is organised into the financial, commercial and legal workstreams, stress-tested from both sides and framed by the sector pack. Each factual statement carries its source, page, provenance and status.
- FDDFinancial due diligence, figures recomputed from source cells
- CDDCommercial due diligence, framed from the pack key drivers
- LDDLegal due diligence, seeded from the regulatory landscape
- Bear caseA forward downside scenario scaled to a named prior crisis
- Bull-lever testThe counterpart to the bear case: the levers that support a long view
- Sector contextNamed competitors and external, cited market reads
Triage · illustrative
DRAFTNorthfield Bearings Ltd
Industrial components · private · WATCH
- Corroborated: Revenue growth, FY25 vs FY24+6.2%
- Sourced: EBITDA margin trajectory18.4%
- Management asserted: Net leverage covenant headroom0.6x
- Needs analyst input: Pending litigation exposureOpen
Illustrative Triage view. Company and figures are examples, not a real target.
One deep read, drawn into three surfaces.
For listed companies the shorter surfaces condense the deeper ones, so a triage cannot say more than the 2-3 pager it comes from, and its citations carry down unchanged.
Triage
1 page · 5 to 10 min · scored by the analyst
Ten labelled facts, five criteria each shown with its evidence for the analyst to score, a band calculated from those scores, and a short summary with open items.
2-3 pager
Up to 5 pages · about 30 min · analyst release
The tear-sheet. The workstreams, the bear case and its counterpart, and a recommendation backed by citations. It is pitched at the level a human analyst would write, not a full transcript.
Investment paper
About 20 pages · about 90 min · analyst release
The investment committee paper: full thesis, modelled scenarios, sector read-across and named kill criteria. It may prompt a deeper pass first, still the same engine over a richer analysis.
Portfolio monitor
A two-page surface for each held position: information, notes, pricing, cap table and manually entered financials.
Chat
Single-turn answers that cite the underlying claims and decline to go beyond the source.
Report formats
Each surface is generated and checked against a stored format contract that your team can edit, so the shape of a report is a house decision rather than the model's.
The differentiator
Industry depth is a generation contract, not a prompt.
A sector is a methodology pack: key drivers, regulatory landscape, named competitors, key metrics and the recurring analytical moves that mark out a good analyst. The engine must attempt a claim for every metric, cover every competitor, and carry out every move, not just name it. Where the sources do not support a pack element, it becomes an explicit open item and is never quietly dropped.
Packs are written in a self-serve editor, versioned like code and recorded against every analysis they produce. Breadth comes from the engine and the authoring path rather than a fixed library.
The pack contract
- A claim attempted for every key metric
- Every named competitor surfaced in context
- The LDD prompt seeded from the regulatory landscape
- The CDD framed from the key drivers
- Every analytical move performed, with real arithmetic
The rules that hold it together
One deep read, many surfaces
The deep read happens once per company. The shorter surfaces are inexpensive projections over claims that have already been extracted.
Depth is data, not a code path
A full analysis and an outline are the same engine working over richer or thinner evidence. There is no second engine to drift out of line.
Thin evidence holds a report back
Public sourcing is assessed once per run. Where the expected documents are missing, every surface from that run stays in draft, whatever its other readings.
The model never does arithmetic
Every figure is recomputed in code from pinned source cells. A figure that cannot be recomputed cannot be marked corroborated.