What a Defensible Underwrite Looks Like: The IC Memo Standard

Every investment committee has a version of the same moment. A memo is on the table, the recommendation is clear, the returns work. Someone asks where a particular number came from.

The good answer is a document name, a page, and a date. The common answer is "it is in the model," or "that came from the broker's pack," or "that is the assumption we normally use." None of those is a source. They are all descriptions of where the number was last seen.

• The standard

  5

  Things every number needs

  Source, location, as-at date, path, and the definition basis

• The leakage

  40%

  Of CAM reconciliations carry material errors

  One service-charge analysis; a route for income leakage nobody watches

• The generic rate

  1 to 6%

  Of spreadsheet formula cells carry an error

  A generic research finding, not a measurement of real estate models

This is not a diligence failure in the usual sense. Nobody has been careless. The number probably is roughly correct. The problem is narrower and more structural: the memo cannot demonstrate why it is correct, which means the committee is being asked to approve a conclusion rather than review an argument.

The standard that resolves this is not "is every number right." No memo can promise that. The standard is that every material number can be traced to something outside the memo, and that the path is short enough to check in the meeting. That is a different discipline from accuracy, and it is the one that survives contact with an LP, an auditor, or a re-underwrite three years later.

The Smart Bricks Research Lab has spent the past several months studying what that discipline actually requires, working from published research on formula-level model auditing and from our own screening work. What follows is the standard as we now apply it.

The first thing that becomes obvious when you try to source every number is that sources are not interchangeable. There is a hierarchy, and the useful discovery is that part of the industry has already written it down.

RICS, in its guidance on comparable evidence in real estate valuation, grades evidence in three categories. Category A is direct transactional evidence, comprehensive and current. Category B is the broader data layer, published indices and databases among it, which points you in a direction without pricing the asset in front of you. Category C is background material. More significant than the grading is the obligation attached to it: the standard obliges members to retain a record of whatever evidence sits behind any figure they report. The duty to show your working already exists inside the valuation discipline.

What does not exist is the same discipline applied to income. No standard-setter publishes an equivalent ranking for the documents a rent roll is built from, which is how the underwriting side of a firm ends up markedly less rigorous than the valuation side about the same asset. Extending the RICS logic from comparables to income evidence gives something like five tiers.

At the top sits the registered or officially filed record: a lease lodged with a public registry, a title record, a filed set of accounts. These are not merely reliable. They are adversarially reliable, in that a counterparty had to accept them and a third party recorded them. Below that sits the executed document held by a party to it, then independent third-party measurement, then the seller-prepared summary, and at the bottom, opinion.

The uncomfortable observation is not that firms use the lower tiers. Sometimes there is nothing else. It is that the tier is not recorded, so a number sourced from a seller's summary and a number sourced from a registered record arrive at the committee looking identical. The memo flattens the hierarchy, and with it the reader's ability to calibrate how much weight to put on any given line.

Where the discipline does exist, it is usually because someone external insisted. Lenders run an explicit test on lease review: whether the lease review form matches the rent roll. That is what an evidence standard looks like once it has teeth, and it exists because a lender required it rather than because the acquirer chose it.

RICS gradeTier of income evidence

1. 1RICS grade A

   Registered record

   Filed with a public registry or statutory body. A counterparty accepted it and a third party recorded it.

   Registered lease, title record, filed accounts

   And underwritten as if from here

2. 2RICS grade A

   Executed document

   Signed and binding, but supplied by a party with an interest in how it is read.

   Lease and amendments, loan agreement

3. 3RICS grade B

   Independent measurement

   Assembled by someone with no position in the deal.

   Transaction evidence, regulated cost benchmarks

4. 4RICS grade B

   Seller-prepared summary

   The most convenient document in any deal, and structurally an argument for a price.

   Rent roll, operating statement, offering memorandum

   Most deals are presented here

5. 5RICS grade C

   Opinion

   No document behind it. Useful for orientation, not for underwriting.

   Broker guidance, verbal confirmation

RICS grades comparable evidence A, B and C, and obliges members to retain a record of whatever evidence sits behind any figure they report. No equivalent standard governs the rent roll, so the tier is rarely recorded and a seller’s summary reaches the committee looking like a registered record.

"Sourced" is doing a lot of work in most processes. In practice a number is only traceable if five things travel with it: the source, named specifically enough that someone else could request it; the location inside that document; the as-at date, because a rent roll is a photograph rather than a fact; the path, meaning whatever happened to the number between the source and the memo; and the definition basis, meaning what the number actually measures.

Naming the document is harder than it sounds, because for most material numbers there is no single document. A tenancy is typically a stack: an original lease, a first amendment, a second amendment restating part of the first in different language, a commencement letter from the tenant's solicitor carrying a date that contradicts both, and a side letter filed somewhere nobody looks that overrides more than anyone remembers. Deciding which instrument governs is a judgment call, and it is made before any figure gets extracted. Get that precedence wrong and everything downstream inherits it: the rent is read off a superseded escalation clause, the expiry comes from a date the commencement letter already replaced.

What makes this failure hard to catch is that the resulting number is still perfectly citable. It has a document, a clause, a date and a path. It is simply pointing at the instrument that lost. So the first part of the test is not only which document a number came from, but which document governs, and whether anyone wrote down why.

The fourth is the one that gets argued about. Almost no material number arrives unchanged. It is annualised, grossed up, net of something, adjusted for a rent-free period, converted, or reconciled against a second source. If the path is not recorded, the number in the memo is not the number in the document, and nobody can say by how much or why.

The fifth is the one most often missing altogether, and it is the one that quietly breaks comparison. Two submarket vacancy rates can each be sourced, located, dated and path-traced and still describe different stock, because the brokerages behind them used different size thresholds and different boundaries. Vacancy and availability are used interchangeably although availability also counts space marketed but not yet empty. A face rent quoted as though it were effective ignores rent-free periods and fit-out contributions. Practitioners working on metric consistency describe drift running along five dimensions at once: names, definitions, values, periods and currency. A committee comparing two deals expressed on different definitions is comparing unlike with unlike, however well cited each side happens to be.

Net operating income, as stated in the memo

$4,120,000

1. Part one

   The source

   Which document governs, not merely which one the figure came from. A tenancy is a stack: amendments, commencement letters and side letters change which instrument wins.

2. Part two

   The location

   Page, clause, schedule row, or sheet and cell. The part almost always omitted.

3. Part three

   The as-at date

   When the fact was true. A rent roll is a photograph, not a fact.

4. Part four

   The path

   Every transformation between source and memo, including hand adjustments.

5. Part five

   The definition basis

   What the number actually measures. Vacancy is not availability; face rent is not effective rent. The one most often missing altogether.

Metric drift runs along five dimensions at once: names, definitions, values, periods and currency. A committee comparing deals on different definitions is comparing unlike with unlike, however well cited each side is.

The five-part test above is a test of format, and that distinction starts to matter as soon as a language model is doing the work.

A model can populate all five fields. Asked for a figure with its provenance, it will name a document, give a clause reference, state a date, and describe the adjustment it applied. The output looks precisely like a traceable number. Every field the test asks for is filled in.

The difficulty is that the fields were composed rather than recorded. The figure and its provenance came out of the same step, and neither is evidence for the other. A clause reference produced by the same process that produced the number is not a citation. It is a second claim, and it fails in the same direction as the first.

There is a straightforward way to tell the two apart, and it is worth running against any system a firm is considering buying. Change one input at the source and re-run. If the provenance was recorded, it moves: the figure changes, the stated path changes with it, and the reference now points at the amended clause. If the provenance was composed, the reference stays plausible and stays wrong, because it was never attached to anything that could move. Recorded provenance is falsifiable. Written provenance is not, and unfalsifiable provenance is decoration.

That has an architectural consequence worth being explicit about. The citation cannot come from whatever writes the prose. It has to fall out of whatever does the arithmetic, as a by-product of having done it: this figure was produced from these inputs by this operation, and each input either sits at a named place in a document or was produced the same way one level down. Walk it far enough and you either reach a document or you reach the point where the chain stopped being derived and started being asserted. Both outcomes are useful. Only one of them is a citation.

So the five-part test is necessary and not sufficient on its own. It needs one property that is not a field at all: the five parts have to be a record of how the number was made rather than an account of it.

Applying the five-part test to real memos surfaces the same three failures repeatedly.

Income that is not income

A rent roll states contracted rent. The executed lease states what was agreed. A bank statement states what was collected. These three numbers are different, they diverge in predictable directions, and they are frequently used interchangeably in the same memo.

The divergence is measurable wherever anyone bothers to measure it. Rent rolls depart from executed leases through misstated expiries, missed escalation steps and wrong areas, and the errors run in the direction of overstating income. Amendments fail to reach the abstract, so figures reflect a base lease that has been superseded. On service charges, one analysis puts material errors in 40% of common area maintenance reconciliations, a route for income leakage nobody is watching. And the risk inside the lease itself often goes unpriced: break clauses appear in 38.5% of UK leases, up from 22.7% in 2007, carrying roughly a 47% probability of vacancy at the break. A break nobody priced leaves a meaningful share of the underwritten income exposed, and it is invisible in a rent roll that reports passing rent alone.

Assumptions with no owner

Every model contains inputs that were chosen rather than observed: rent growth, exit yield, downtime, capital expenditure allowance. Each is defensible. Few are attributed. Estimated rental values and growth assumptions are routinely entered with no cited source at all, or taken from a single market average.

There is a documented reason this matters more than it appears. Studies of appraisal overvaluation find that analysts over-weight favourable comparables and adjust asymmetrically once they already know the target price. The bias needs nobody to act in bad faith. It needs only that the number to be justified is known before the evidence is assembled. Selecting comparables to reach a predetermined figure breaches professional standards, which tells you the practice is common enough to have needed a rule against it. An unattributed assumption is where that pressure lands, because nothing in the memo records who chose it or what they were looking at.

Derived numbers with no path

The most common single defect. A figure is calculated in the model, exported, adjusted by hand in the memo for some sensible reason, and reported. The adjustment is not recorded. Six months later nobody can reproduce the number, including the person who made it.

There is a second, separate question, and conflating it with sourcing is a mistake we made early in our own work.

It is worth first noting how little the industry measures here. There is no standard, continuous indicator of underwriting-model quality. Nothing in the governance stack catches a modelling error before diligence does. The nearest generic benchmark comes from outside real estate: a long-standing body of spreadsheet research puts the share of formula cells carrying an error somewhere between one and six percent, and it is worth being explicit that this is a generic finding about spreadsheets rather than a measurement of real estate models specifically.

Two disciplines do exist, and both are worth knowing about before commissioning anything new. Independent model audit, where an outside specialist works through the model against a formal checklist, is long established; one firm in that business describes a method it has been refining for three decades. And on convention rather than audit, the most widely used real-estate-specific modelling standard sets out a font-colour input taxonomy in which required inputs are visually distinct and owned by the analyst, calculations and cross-sheet links are never overtyped, overrides are flagged, every deal starts from a fresh template, and circular references are avoided precisely because they obscure the audit trail.

A memo can cite everything correctly and still be wrong, because the model that produced its outputs does not compute what the memo claims. Published research on formula-level auditing of institutional models makes this concrete. In one documented framework, models are tested against 99 individual checks across 12 evaluation categories, each weighted across three severity tiers so that a critical failure counts for more than a cosmetic one: an internal rate of return formula pointing at the wrong cash flow row, a bridge balance that stops rolling forward, replacement reserves counted once in operating expenses and again below the net operating income line.

The instructive result from that literature is not the failure rate. It is a case in which a model scored 97.8% against the framework and was still flagged unreliable, and the reason is worth stating precisely: the evaluator's own definition of correctness did not fully cover the territory. On opportunistic deals, where the cash flow behaves differently from one phase of the hold to the next, a fixed set of checks develops blind spots, and a model can converge on a high score while the thing evaluating it still cannot vouch for it. The published work names this the difference between a defined problem and a well-defined one. A high score on a well-cited model is not the same as a model you can trust on the next asset.

A memo is not defensible because its numbers are right. It is defensible because a stranger can retrace them.

Smart Bricks Research Lab

So the standard has two halves: whether every material number can be traced to something outside the memo, and whether the thing that produced it computes what the memo claims. Firms tend to have a process for neither. The ones that have a process usually have it for the first only.

Stated as something an analyst can run before a memo goes out.

On every material number

• Source document named specifically, not by category
• Where a document stack exists, the governing instrument is identified
• Location given: page, clause, row, or sheet and cell
• As-at date stated
• Every transformation recorded, including hand adjustments
• Evidence tier visible so the reader can weight it
• The definition basis is stated, so two figures compare safely

On every assumption

• Labelled as an assumption, not an observation
• The person who selected it is identified
• The evidence or convention behind it is stated
• Any departure from the firm default is explained

On the model behind it

• Return metrics reference the intended cash flow rows
• Debt balances roll forward and reconcile every period
• Reserves and capital items appear exactly once
• The deal started from a fresh template, not a marked-up copy
• No calculation or cross-sheet link overtyped; overrides visible
• No circular references, which obscure the audit trail
• Tested by something other than what built it
• Formulas still correct if the cash flow pattern changes

On the memo as a whole

• Any single number checkable in under a minute
• Nothing material rests on a source that cannot be produced

Nothing on this list is technically difficult. All of it is tedious, which is why it is not done consistently by people, and why it is well suited to systems that do not get bored.

The objection to this standard is that it is more work, and at the level of a single memo that is true.

The relevant comparison is not against a memo without citations. It is against the cost of the alternative, which is paid later and in worse currency. A rent roll that does not survive diligence is one of the recognised reasons a buyer reopens price, so the mistake gets paid for either in the closing number or in a renegotiation halfway through. At its extreme end, a misstated rent roll has produced collapsed transactions, deposits kept by the seller, and claims for rescission or damages; missed lease provisions are established grounds for a professional negligence claim against an adviser. Data-quality practice has a rough shape for this that is worth borrowing even though it is folklore rather than accounting: about a unit of cost to prevent an error, ten to correct it, a hundred once it has caused a failure.

It is also worth noticing who eventually asks. The committee asks under time pressure and wants the argument to hold long enough to decide. An auditor asks later, and wants the same figure to come out the same way twice. A limited partner asks years afterwards, and wants a number reported in a past quarter tied to something that still exists. Those are three different interrogations, and a memo built to this standard answers all three out of one structure. That is the difference between doing the sourcing work once and doing it three times.

There is also a compounding effect that is easy to miss. A memo built to this standard is reusable. The sourcing work done at acquisition is the same work asset management needs at every subsequent reporting date, and the same work the next buyer's diligence will ask for at exit. Done once and recorded, it is an asset. Done informally and forgotten, it is repurchased every time somebody asks.

Which is the practical reason the standard is worth adopting before it is required rather than after. The firms we see moving fastest did not begin with a technology programme. They began by picking one memo and refusing to let any number into it that could not be traced, then finding out how much of their existing process could not meet that test.

The memo is not the paperwork around the decision. In an institutional process, the memo is the product. It is worth building like one.