On the memorandum.

The first product decision we made — before any code was written, before the company had a name on the wall — was that the artefact would be a memorandum.

It wasn't an obvious decision. A chat transcript is the easiest thing a diligence tool can produce. Customer asks, model answers, transcript is the deliverable. The transcript reads as if the work has been done. It can be rerun on a different model with no engineering effort. It ships in a day.

We don't ship one, because what a partner ends up signing isn't a transcript. It's a document. The reason the difference matters is buried somewhere most software vendors haven't looked, which is what the partner is actually being asked to defend.

A partner who signs a diligence memorandum is committing the House to a position. The position will be read by an IC, by a co-investor, by counsel on a structuring call, by an auditor at year-end, and — when the position turns out to be wrong, which on a long enough timeline it will — by a plaintiff's lawyer.

The partner can't defend the position by saying the model produced it. The defence is structural. It looks like this: here is the question we asked, here are the materials we read, here is the claim we made, here is the span in the source that supports it, here is the reviewer who agreed, here is the version that was locked, here is when the lock happened, here is the audit log that records all of it.

A chat transcript has the question and the answer. A memorandum has the question, the answer, the citation, the refusal, the reviewer, the version, the timestamp, and the log. The difference isn't a polish difference. It's a difference in what's available to the partner the morning after.

Four things, in the rough order they're missing from the diligence products I've evaluated.

First, span-level citations. Every claim the memorandum makes carries an inline anchor back to a span in the source: filename, page, paragraph. A claim that can't find its span doesn't survive the validator; the engine drops it before the memorandum reaches the reviewer. The reviewer sees the dropped claim and decides what to do. The reader of the locked version doesn't see speculation, because speculation didn't survive the write.

Second, refusals that read like the partner wrote them. Where the materials don't support a claim that a confident model would otherwise produce, the schema requires a short reason in a refused field, written in the partner's vocabulary — not the model's. The phrase "I cannot determine" doesn't get into the schema. The vocabulary we allow is finite, and it's drawn from how partners refuse claims in memoranda they wrote by hand. A reviewer trained on it learns to read the absence of a sentence as a finding.

Third, structure that survives the model. The memorandum is rendered against a schema: sections, fields, citation arrays, refusal sentinels. The schema lives in the same repository as this letter, version controlled, with an ADR for every meaningful change to it. The model writes against the schema. The schema does not write against the model. When the model changes — which happens often enough that we have a process for it — the schema doesn't. The committee that read last quarter's memorandum reads this quarter's the same way.

Fourth, a version trail. A locked memorandum is immutable. A new version is a new row, not an overwrite. The trail records which prompts ran, which model produced each section, which reviewer locked when, and what changed between locks. A general counsel can read it. A regulator can replay it. A successor partner, three quarters from now, can find the moment a claim was added and the moment it was checked.

None of those is a marketing feature. They're the engineering that lets a House sign and defend an output, which is the work the engagement was paid for.

The shape of the artefact is easier to show than to describe. The block below is from a memorandum the engine produced in a recent rehearsal — the same chrome, the same citation pills, the same refusal posture you'd see inside a live engagement.

Manager IDD · §3.2 · Where we disagree with the lead

Operational improvements are partially supported.

The lead memorandum prices the base case at a 2.1× MOIC over five years on the strength of three operational improvements: pricing discipline [S-014 · p.42 · ¶3]LeadMemo-2026Q1.pdfPricing actions taken in 2025 increased ASP by 4.1% across the core SKU set., route density [S-014 · p.47 · ¶1]LeadMemo-2026Q1.pdfRoute density improvements yielded a 220 bp reduction in cost-to-serve., and SG&A consolidation [S-018 · p.11 · ¶4]Sponsor-OperatingPlan.xlsxTargeted SG&A reductions of $14M annualised by FY27, sourced from regional office consolidation..

The first two are supported by primary materials in the data room. The pricing action is visible in the company's most recent rate cards [S-022 · p.3 · ¶2]2026-Rate-Card-Schedule.pdfEffective rate increases of 3.8–4.4% by SKU family, dated 1 January 2026., and the route-density claim is supported by the operating company's own internal density study [S-029 · p.14 · ¶5]Density-Study-Q3-2025.pdfCost-to-serve reductions of 210 bp realised in pilot regions; 220 bp the run-rate projection at full deployment..

The third — SG&A consolidation — is supported only by the sponsor's own operating plan. The data room does not contain the regional headcount detail, the office-lease assumptions, or the severance reserve required to verify a $14M annualised reduction. The base-case MOIC, recomputed without the SG&A line, falls to 1.9×.

What I'd want a reader to notice in the block isn't long. It's three citations against the data room, one absence written in plain English, and a recomputed number that drops because the supporting evidence isn't there. The block doesn't say which sponsor, which company, which region; those are inside the chamber and don't belong on a marketing surface.

A reviewer reading the block opens each cited source and checks the span in about fifteen seconds — the citation pills hover with the excerpt, and each hover gets logged. The committee reading the locked version sees the same block, the same citations, and the same absence, with a base case that has been recomputed on the page rather than buried in a footnote.

Evidence ledger · Manager IDD · §3.2

Base case · operational improvements

ClaimStatusSourcePricing disciplineSupportedS-022 · rate cardsRoute densitySupportedS-029 · density studySG&A consolidationUnsupportedSponsor plan only

Base-case MOIC

2.1×

A chat transcript is cheap. A memorandum is not, and the cost is worth itemising, because it's what the House is actually paying for.

There's a retrieval cost. Every claim has to find its span. The retrieval stack — BM25, vector search, fusion, reranker — runs against an indexed chamber and returns only the spans the model is allowed to cite. A claim whose retrieval returns nothing is a claim the validator won't allow into the memorandum. The cost is the index, the embeddings, the run-time retrieval calls. The benefit is that no claim gets out without a citation, which is what makes the document defensible after the fact.

There's a schema cost. The model's output is parsed against the memorandum schema before the document is written. Fields that fail get coerced or refused. Fields that pass enter the document. The cost is the validator, the coercion layer, and the audit log that records every coercion — the part of the engine that most resembles compiler work, and the part where most of the engineering hours have gone. The benefit is that the document a partner signs is structurally the document the schema describes, every time.

And there's a review cost. A reviewer reads the memorandum claim by claim before it's locked. The reviewer can drop a claim, change a citation, request a refusal, or accept the section as written. The lock is the partner's signature; until it lands, the memorandum is a draft. That review time is the cost most clients underestimate when they first see the price. It's also the cost that makes the difference between a memorandum and a generated artefact: somebody whose name will be on the document has read every claim in it.

The total is meaningfully higher than a chat transcript and meaningfully lower than producing the same memorandum by hand. What the House is buying is the difference between the second number and the third — the document it couldn't have written at the same speed, against the same materials, with the same evidentiary discipline.

A chat surface is useful. We use one inside the chamber for the reviewer — a partner asks the engine a follow-up about a section, sees the answer with its citations, decides whether to fold the answer into the memorandum. The chat is a tool the partner uses on the way to the document. It isn't the document.

The reason is the asymmetry of consequences. A chat answer that turns out to be wrong is a conversation the partner had with a tool. A memorandum claim that turns out to be wrong is a position the House took. The committee, the auditor, the regulator, and the LP all read memoranda. None of them read chat transcripts. The artefact the House gets judged by is the one with the signature on it.

We build that one.

If you're a principal evaluating us for the first time, the practical consequence is this. On the first read of our output, the memorandum will feel slower than a chat-shaped product. The engine declines to write sentences a reviewer would have written without thinking. It softens claims a reviewer would have softened anyway. It puts a citation under every assertion, and the citations stay visible on the page.

On the second read, the memorandum will look like the ones the partner has always written, and the audit log will be the one a general counsel has been asking for since the first quarter of 2023 and has never received from a vendor. That's the artefact. The House signs it. We've built our way toward it.

— A.O.