What an underwriter can know, and cannot.

There are three kinds of claim in a diligence memorandum, and they get treated the same way by everyone except the partner who has to defend the memorandum eighteen months later. The first kind you can verify against a single source. The second you can infer from a small chain of sources. The third is conjecture, and a careful person doesn't write it down.

Most of what makes a diligence practice work — what makes it survive being read after the deal goes sideways — is the line between the second and the third. It's not a line you can hold by reading carefully. It has to be held by the scaffolding underneath the writing.

This letter is about where that scaffolding sits in our engine, and what changes about a memorandum when it's there.

A verifiable claim has one source in the data room, and the source says exactly what the memorandum says.

The covenant in the credit agreement requires a fixed-charge coverage ratio of at least 1.25× as of quarter-end. The compliance certificate from last quarter, signed by the CFO, reports a ratio of 0.96×. The covenant is breached. A reviewer opens the credit agreement, finds the covenant, opens the certificate, finds the calculation, and verifies the arithmetic in about fifteen seconds. The memorandum cites both documents in the same paragraph, by span. Anyone on the committee can do the same.

The sponsor's most recent quarterly shows $48M of cash on the balance sheet. The lender's most recent borrowing-base certificate reports $42M of revolver availability. Combined liquidity, $90M. The memorandum says so. Verifiable.

This is the engine's natural register and the part I'm least worried about. The discipline is to keep the claim in the language the source used, cite by span, and refuse to extrapolate. The temptation in any system that produces prose is to nudge a claim into something more confident than the citation supports. "The covenant is breached" wants to become "the company is in material breach of its credit agreement." Those are not the same sentence. The longer one requires reading the credit agreement's definition of material breach, checking the cure provisions, and confirming that no notice has intervened — a chain of work the citation alone doesn't carry. Our engine writes the shorter sentence.

An inferable claim has no single source that says it on its own, but a small number of sources, taken together, say it. The chain has to be short enough that a reviewer can follow it.

A real example. The borrower's audited statements show EBITDA margins declining year over year for four years. The industry peer set — three named comparables, drawn from the most recent filings — shows margins rising over the same four years. The borrower's most recent operating review describes the loss of a major customer in the second half of the most recent fiscal year. The claim that can be drawn from those three sources is that the company is losing share in a growing industry, and that the customer loss in the most recent half is consistent with that pattern rather than a one-off event.

The chain has three links. Each link is verifiable on its own. The claim is the conjunction. A memorandum that makes the inferable claim has to show the conjunction in the same paragraph — three citations, each anchored to its source — and write the inference in the underwriter's register. "The company appears to be losing share." Not "the company is losing share." The conditional verb is the discipline. The inference is short enough to follow, and the verb tense is honest about it being an inference.

The trouble with inferable claims is that they look on the page like verifiable ones. A reader doesn't always notice the assertion is the conjunction of three sources rather than the citation of one. A reviewer in a hurry sees three citations, accepts the claim, doesn't open any of them. The engine can't fix the reviewer's hurry. What it can fix is the verb tense. "Is losing" is a fact. "Appears to be losing" is an inference. The schema enforces the distinction at the field level, because in my experience it isn't always enforced by the writer.

A conjectural claim is one where the sources, taken together, don't support the claim, but the claim could be true, would be useful if true, and is one the model will produce if you ask the right way.

These are the dangerous ones. They sound plausible. They flatter whoever wanted the claim to begin with. They survive a first reading because they read as if they were inferable. They fall apart when a reviewer goes looking for the chain, finds nothing, and either pulls the claim or — far more often — runs out of time to pull it and leaves it in.

The example I've argued about most in my own practice is sponsor competence. A memorandum will sometimes read something like: "The sponsor has a strong track record of operational improvement in middle-market industrials." That line sits unremarked in dozens of memoranda I've read. It is almost never supported. The data room contains the sponsor's marketing book, which says so, and the marketing book isn't a citation. It's an assertion the sponsor would like the reader to repeat. A real underwriting of sponsor competence requires returns by fund, by vintage, and by deal, decomposed into entry-multiple expansion, financial leverage, and operational improvement — and the data for the last category is, in most rooms, not in the room.

Our engine won't write the operational-improvement sentence. It writes, in the same paragraph: "The sponsor has marketed itself as operationally focused. The data room does not contain the deal-level attribution required to verify operational alpha." The committee gets the same information. It also gets the absence of the supporting data, in plain words. A reviewer can decide to go ask the sponsor for the missing attribution. If the sponsor provides it, the conjectural claim becomes inferable. If the sponsor declines, the absence is itself a finding.

That's what we mean when we say the engine refuses. The phrase "I cannot determine" doesn't appear in the schema. The model isn't allowed to write it. What is allowed is the partner's own vocabulary for what the materials don't contain. A few engagements in, partners stop reading the refusals as gaps in the engine's coverage and start reading them as findings, which is how the trained eye in this business already works on the read side.

Three kinds of claim

• Verifiable

  The covenant is breached.

  S-014 · §3.2

• Inferable

  The company appears to be losing share.

  S-018 · p.3S-022 · p.8S-029 · p.14

• Conjectural

  The sponsor has a strong operational track record.

  Refused

  Materials do not contain the deal-level attribution required to verify operational alpha.

A diligence practice that doesn't enforce the line between inferable and conjectural produces memoranda that look defensible on a first read. Every sentence has a citation. Every claim sounds careful. The document reads well at first reading. And it contains three or four sentences the materials don't support.

Those three or four sentences are the ones that come back.

They come back in a quarterly review when the committee has to explain why an asset underwritten as "well-positioned in a structurally growing segment" is being written down. They come back in an LP letter, when an outside director asks why the memorandum didn't flag the customer concentration. They come back in a fund wind-down, when the LP sues and the memorandum becomes a document that has to be defended sentence by sentence in front of counsel.

The cost of refusing the conjectural claim, on the day of underwriting, is small. A partner who wanted the deal to clear is mildly frustrated. The IC meeting runs fifteen minutes longer because the absence has to be discussed. One of the sponsor's marketing assertions doesn't make it into the recommendation. The cost, eighteen months later, of having included a sentence the materials didn't support, is much larger, and it falls on the partner who signed, not on the engine that wrote it.

A diligence practice that internalises the trade-off writes shorter memoranda, refuses more often, and ends up trusted more by the committee that reads what it produces. That's what we've been building toward.

The most common question I get from Houses evaluating us for the first time is whether the engine's refusals are a feature or a bug. It's a fair question. A reviewer who has read a hundred memoranda from a sell-side advisor — every one of which is confident, every one of which makes the sponsor competence claim, every one of which is read by a committee in twenty minutes — finds it strange to see, in our output, a section that quietly declines to make the claim.

It's a feature, and it's the most important one the engine has. The model could write the confident sentence. The reviewer might prefer it in the first week. But the memoranda I've written by hand, over the years of underwriting that informed the design of the system, are the ones that refused to make the claim the materials didn't support. Those are the memoranda I would still defend.

The engine writes that kind of memorandum because that's the kind a careful underwriter would write. The vocabulary, the schema, the citation discipline, the refusal sentinel — that's the engineering underneath the posture. The model is the writer. The posture is the work, and the work is what we've spent two years on.

— A.O.