Why Generic Due Diligence Fails on Private Real Estate: The Register Was Never Asked

An office building changes hands. Counsel has issued a certificate of title with a schedule of registered leases attached. The model has been rebuilt twice during diligence as tenancy documents arrived, and it ties to the seller’s rent roll to the pound. The surveyor has walked the building and put a band around the roof. The valuer has confirmed the price against five comparables. Eleven months after completion the asset manager brings forward a refurbishment of the second floor, timed to the tenant’s expiry in year five, and the lender’s counsel points out that the tenant’s lease is registered with a term running to year twelve; the rent roll had carried the break date as the expiry. The reversion the model priced in year five moves seven years out. The lease that says so had been on the register throughout, listed in a schedule to the certificate that nobody read against the rent roll.

Nothing in that sequence was careless. Each of the four instruments did what its scope required, and each would pass an audit of its own file. The failure sits in the design of the process. The institutional diligence playbook, legal and title review, financial model, physical and technical inspection, broker opinion of value, was drawn up for markets in which the only comprehensive account of an asset was the one the seller assembled. In such a market the pack is the record, and the rational thing to do with it is to test it for internal consistency: does the rent roll tie to the leases in the data room, does the model tie to the rent roll, does the price sit inside the comparables. Every instrument in the playbook is a test of that kind.

The markets where most institutional capital now transacts are not of that kind. The 2026 edition of the Global Real Estate Transparency Index scores 88 countries on 260 measures and places 13 of them in its highest tier, with the United Kingdom first, the United States fourth, the Netherlands fifth and Sweden ninth; those highly transparent markets take more than 80% of direct real estate investment. In the Netherlands, in England and Wales, across the Nordic countries and in the American county system, an independent record exists for who owns an asset, what is charged against it and, in most of them, what was last paid for it. In those markets the pack is a claim about a record that can be read, and the playbook still treats it as the record.

An earlier note in this series set out what a cited figure must carry, and a second described the strata a citation points into; this one concerns the order in which the instruments are run, and what happens when the register is asked last. The Smart Bricks Research Lab is working through that question against the registries’ own published statistics, the standards that define each instrument’s scope, and the academic measurements of what survives a standard process. What follows takes each instrument in turn, sets out what the register holds, shows that the resulting confirmation is measurable and that a horizontal language model fails in the same shape, and proposes a sequence that puts the register first.

Take the four instruments in turn and ask one question of each: what is its working input, and who supplied it.

Legal and title review. Counsel receives the seller’s title pack and the seller’s replies to standard pre-contract enquiries, and requests official copies of the register entries. That is the one point in a standard process at which the independent record is read. The output is a certificate of title addressed to the buyer and its lender, with schedules of what the register disclosed: registered leases, charges, easements, covenants. The certificate records what was found. It does not record what the rent roll says about the same tenancies, because counsel is not given the rent roll and is not instructed to reconcile it. The register’s coverage explains part of this. In England and Wales a lease is registrable only where more than seven years of its original term remain unexpired, so most commercial tenancies by count, and the rent under all of them, sit in documents the register does not hold. For most income there is nothing to take it from except the seller’s word, which is a reason to label the rent roll as the seller’s and no reason to leave the registered leases unreconciled against it.

The gap between a registered fact and a diligence process was drawn sharply by the UK Supreme Court in 2020. A developer had built 23 affordable homes, 13 of them on land burdened by a registered covenant against building, knowing of the covenant. The court refused to modify the covenant and described the conduct as a cynical breach. The lesson for an acquirer is narrower than the court’s finding: the covenant was on the register for the whole period, and knowing it existed was a different act from testing the intended use against it.

The financial model. The model is built from the rent roll in the offering memorandum and the seller’s operating statements. As tenancy documents arrive the roll is corrected, an expiry here and a rent-free period there, and the corrected roll is what the committee sees. What the model never receives is the identifier the register uses. No title number, cadastral object or parcel number sits on any row, so there is no line along which a registered fact could be joined to a model input. The arithmetic is tested at length. The premise the arithmetic runs on remains the seller’s.

Physical and technical diligence. The scope here is set by the standard itself. ASTM E2018-24 defines the walk-through survey at the heart of a baseline property condition assessment as nonintrusive observation of readily accessible and observable improvements, and defines the opinions of cost it produces as preliminary, order-of-magnitude budgetary projections. Those are honest definitions, and a baseline assessment claims nothing beyond them: the condition of a building on one day, seen from where a person can stand, priced to a band. Address and building registers in the Netherlands and Sweden hold construction year, use and area for every building in the country, and energy certificate registers hold performance ratings; a standard process compares almost none of it to the survey.

The broker opinion of value. Comparables are selected, and the party selecting them holds a position in the outcome. Valuation standards prohibit choosing comparables to arrive at a figure already known, and the prohibition exists because the pull towards doing so is constant rather than occasional. In the four jurisdictions discussed here a registered or public price exists for many of the comparables an agent chose and for many more that the agent did not, and a standard process consults those prices, where it consults them at all, as a cross-check after the number has been settled.

Each instrument, then, runs on an input the seller supplied or framed, and each tests the pack for consistency with itself. Exhibit 1 sets the four against the independent record that exists for each.

Instrument	Working input	What it tests	Independent record available	Joined to the model?
Legal and title review	Working inputSeller's title pack, replies to standard enquiries, data room	What it testsWhether the seller's title story is consistent with the extract counsel pulls	Independent record availableTitle and charges registers, registered leases (E&W over seven years), company filings	Joined to the model?No; lands in a certificate of title, not in the model
Financial model	Working inputSeller's rent roll, offering memorandum, operating statements	What it testsThe arithmetic of the seller's income story	Independent record availableRegistered deeds and mortgages, tenants' filed accounts, registered prices where published	Joined to the model?Rarely; the rent roll is keyed in and the title number is not
Physical and technical DD	Working inputNonintrusive walk-through of readily accessible components (ASTM E2018-24, 3.2.65)	What it testsVisible condition on one day, costed to order of magnitude (3.2.35)	Independent record availableBuilding and address registers (BAG; Swedish building part), energy certificate registers	Joined to the model?Rarely; the survey is a PDF, the register is not consulted
Broker opinion of value	Working inputComparables selected by an agent with a position	What it testsThe price the agent believes the market will accept	Independent record availablePublic price registers (Kadaster, Lantmäteriet, NLS Finland) and recorded deeds (US counties)	Joined to the model?Sometimes, as a cross-check after the number is chosen

Each instrument runs on an input the seller supplied or framed, so it tests the pack for internal consistency. An independent record exists for most of the facts, and in a standard process it is read once, by counsel, and never joined to the model.

Source: Smart Bricks Research Lab; ASTM E2018-24 definitions 3.2.35 and 3.2.65; HM Land Registry Practice Guide 25; Kadaster; Lantmäteriet; National Land Survey of Finland

The independent record differs by jurisdiction in what it holds, who guarantees it and how it is priced. Two things are constant: the register knows who owns the asset and what is charged against it, and the register does not know the rent.

England and Wales. HM Land Registry guarantees the title it records, keeps a charges register of mortgages and other encumbrances against each title, and registers leases with more than seven years of their original term unexpired. Its dataset of titles held by UK companies and corporate bodies runs to approximately 3.2 million records and is refreshed monthly, so the registered owner of a commercial asset held through a company can be matched to a title number without asking the seller. The corporate side of that match is public too: the Companies House register held 5,427,787 companies at the end of the 2025 financial year, each with filed accounts and officers, so a tenant covenant or a selling vehicle can be checked against something other than the offering memorandum. Price is the register’s weak point for a commercial buyer: the open price series is residential by definition, and commercial consideration appears only where the title itself states it.

The Netherlands. Kadaster registered 517,500 deeds and 456,500 mortgage documents in 2025, and answered 29.0 million online information requests in the same year, a volume that says how routinely the record is consulted by the professions, and how rarely by the acquirer’s model. A copy of the registered deed for any parcel can be ordered, and the deed carries the consideration. The published price product is narrower: purchase price, sale date and plot area for houses and apartments bought by private individuals since 1 January 2003, available to anyone for a few euros. Beside the cadastre sits the BAG, the national register of addresses and buildings, maintained by municipalities and published as open data, which gives construction year, use and floor area for every building in the country.

Sweden and the wider Nordics. Lantmäteriet’s real property register holds ownership and mortgages for about 3.2 million properties in its land registration part, updated every weekday, and describes about seven million buildings across all 290 municipalities in its building part. The property price register, recording price, date, buyer and seller, is public. Finland goes further in statute: the Property Purchase Price Register is a public register under the Act on the Property Purchase Price Register of 1980, compiled from the notices of public purchase witnesses and recording the object, the price and the parties. Norway’s land register, the grunnbok, is described by its registrar as a public register of registered rights and encumbrances in real property. Denmark’s digital land register shows the owner, the assessed value, security interests, easements and covenants for any property without a login, and is run by the courts. None of the five records commercial leases or the rent under them.

The United States. The United States records deeds and mortgages county by county, across 3,069 county governments, and the record guarantees no title; that is why a title insurance industry exists. The record is fragmented, and it is real. Recorded deeds, recorded mortgages and liens, and in many states a stated consideration or a transfer tax from which one can be derived, are available for every parcel through the county recorder, and 2,738 recording jurisdictions now accept electronic recording. A process that treats the American record as unusable because it is unguaranteed has confused two properties of a register: whether it is warranted, and whether it can be read.

The pattern is the same in every jurisdiction. Ownership, encumbrance and, in most of these markets, price are matters of record; rent, most commercial lease terms and every forward assumption are not. A standard process takes the first group on the seller’s word along with the second, when only the second requires it.

Jurisdiction	Register	Ownership	Encumbrances	Price	Commercial leases	Scale and access
England and Wales	RegisterHM Land Registry	OwnershipYes, title guaranteed	EncumbrancesCharges register	PriceResidential Price Paid Data; commercial consideration in the title where stated	Commercial leasesRegistrable only over seven years	Scale and access3.2m company-held titles published monthly; searchable by title number
Netherlands	RegisterKadaster, with BAG	OwnershipYes	EncumbrancesMortgages registered	PriceDeed of any parcel available; residential price product since 2003	Commercial leasesNot registered	Scale and access517,500 deeds registered in 2025; 29.0m online requests
Sweden	RegisterLantmäteriet	OwnershipYes (lagfart)	EncumbrancesMortgages (inteckningar)	PricePublic property price register	Commercial leasesNot registered	Scale and accessAbout 3.2m properties, 7m buildings; updated every weekday
Finland	RegisterNational Land Survey	OwnershipYes	EncumbrancesMortgages	PricePublic purchase price register (Act 552/1980)	Commercial leasesNot registered	Scale and accessPublic by statute; data via Land Information Service
Norway	RegisterKartverket grunnboken	OwnershipYes	EncumbrancesRegistered rights and encumbrances	PriceRegistered with the transfer	Commercial leasesNot registered	Scale and accessDescribed by Kartverket as a public register
Denmark	RegisterTinglysningsretten tingbogen	OwnershipYes	EncumbrancesSecurity interests, easements	PriceProperty value shown	Commercial leasesNot registered	Scale and accessOwner, value and charges visible without login
United States	RegisterCounty recorders	OwnershipRecorded deeds, no title guarantee	EncumbrancesRecorded mortgages and liens	PriceConsideration or transfer tax, varies by state	Commercial leasesNot recorded	Scale and access3,069 county governments; 2,738 accept eRecording

The register knows who owns the asset, what is charged against it and, in most of these markets, what was paid. It does not know the rent, and in no market does it hold most commercial leases. The US records deeds county by county and guarantees no title; it is fragmented, not absent.

Source: HM Land Registry (PG25; UK companies dataset via data.gov.uk); Kadaster Jaarverslag 2025 and Koopsominformatie; Lantmäteriet; National Land Survey of Finland; Kartverket via data.norge.no; lifeindenmark.borger.dk; NACo; PRIA

If the four instruments confirm the pack rather than test the asset, seller-framed numbers should survive institutional diligence at scale, and they should survive in the direction the seller prefers. Both predictions have been measured.

Griffin and Priest, writing in the Journal of Finance in 2023, compared the income underwritten on commercial mortgage-backed securities loans against the income the properties later reported. On 29% of loans, underwritten income exceeded actual net operating income by 5% or more. The overstatement varied persistently by originator, was predictable before the loan was made, and predicted later distress. Every loan in the sample had been through an appraisal, a lender’s underwriting model and a title review, and the income figure the process produced was still, on more than a quarter of loans, materially above what the building went on to earn. The instruments ran, and the number they ran on was the seller’s.

The fund-level evidence points the same way. MSCI measured 145 US closed-end real estate funds of the 2005 to 2007 vintages, the years in which packs priced at the peak of the cycle passed through diligence in volume, against its US quarterly property index, and found that 72% underperformed on both an unlevered and a levered basis. Diligence at the peak was no thinner than at other times; it confirmed the packs it was given, and the packs were priced for a market that was ending.

The process that produces the memo is also universal and slow enough that its content is set by what is available to fill it. A survey of 54 institutional investors by Addepar and the Stanford Initiative on Long-Term Investing, run in June 2022, found that 97% ran a formal investment memo process and 78% required the memo to be approved before capital was committed. The full write-and-review cycle averaged two months, within a range of two weeks to a year. The same researchers found no peer-reviewed article on the investment memo on Google Scholar as of March 2023. A document that nearly every institution produces, that gates nearly every commitment, and that takes two months to write has never been studied; and nothing in the two months obliges the analyst writing it to read the register before the rent roll. The sequence is optimised to assemble the pack, and the pack belongs to the seller.

• 29%

  CMBS loans whose underwritten income exceeded actual property income by 5% or more, after appraisal, model and title review

  Griffin and Priest, Journal of Finance, 2023

• 72%

  US closed-end real estate funds of 2005 to 2007 vintage that underperformed the index, unlevered and levered, 145-fund sample

  MSCI, August 2022

• 97% and 78%

  Investors with a formal memo process, and firms requiring memo approval before investing; average cycle two months; 54 organisations

  Addepar and Stanford Long-Term Investing, 2022 survey

• 58% to 88%

  Hallucination rate of public LLMs on verifiable questions about federal cases; models often accept users' false premises

  Dahl, Magesh, Suzgun and Ho, Journal of Legal Analysis, 2024

• 17% to 33%

  Hallucination rate of retrieval-grounded legal research tools

  Magesh et al., 2024

• 44%

  Precision at 80% recall for the best 2021 model on the CUAD expert-annotated contract benchmark (510 contracts, 41 clause types)

  Hendrycks et al., NeurIPS 2021

Seller-framed numbers survive institutional diligence at scale, and language models reading documents alone accept the premise they are handed. Both failures are measured; neither is fixed by running the same process faster.

Source: Griffin and Priest (2023); MSCI (2022); Addepar and Stanford Long-Term Investing (2022); Dahl et al. (2024); Magesh et al. (2024); Hendrycks et al. (2021)

Diligence that starts from the seller’s pack can only end by agreeing with it. The register has to be read first.

Smart Bricks Research Lab

A horizontal language model applied to a data room does what a pack-first process does, for the same structural reason: its working input is the set of documents it is handed, and it has no second record to hold them against. The evidence for this comes from law rather than from real estate, and it is precise.

Dahl, Magesh, Suzgun and Ho, in the Journal of Legal Analysis in 2024, put verifiable questions about US federal cases to public language models and found hallucination rates of at least 58% for the strongest model tested and up to 88% for the weakest. The more relevant finding for a diligence process is the second one: the models often uncritically accepted the incorrect legal assumptions embedded in the questions put to them. A model asked to confirm a premise confirms it. That is the behaviour of a process that starts from the pack, reproduced in software.

Grounding the model in documents reduces the failure without removing it. Magesh and colleagues tested the retrieval-augmented legal research tools sold by the two largest legal publishers, systems that retrieve authoritative documents before answering, and found that they still hallucinated between 17% and 33% of the time. The point for real estate is uncomfortable. Grounding a model in the seller’s data room grounds it in the seller’s premise; retrieval improves the citations, and the citations point at the pack.

Nor is the reading of the documents themselves reliably solved. The CUAD benchmark, published in 2021, contains 510 commercial contracts with more than 13,000 expert annotations across 41 clause categories, the categories a lawyer reviews in a transaction. The best model tested reached 44.0% precision at 80% recall: tuned to find four in five of the governing clauses, it returned more wrong clauses than right ones. Those are 2021 baselines and models have moved since. The benchmark still shows that locating the clause that governs, inside documents that amend one another, is hard for a system reading text alone, before any question of whether the clause is true.

What follows for a deal team is a design requirement rather than a procurement preference. A model reading the seller’s documents inherits the seller’s premise, and a larger model inherits it more fluently. Structural reading requires two things the documents cannot supply: a rule for which instrument governs when the documents disagree, and a record outside the documents against which the result is tested. The first is a matter of precedence within the pack. The second is the register, and it is a question of what data the system is given and in what order, not of how many parameters it has.

The practical change is small and it is a change of sequence. The four instruments stay, and each keeps its scope. What changes is what they are run against.

Pull the register before the pack. On day one, before the data room is opened, the team orders the register extract for every title, parcel or cadastral object the asset comprises: title and charges, registered leases in England and Wales, deed and mortgage registrations in the Netherlands, Sweden and Finland, the recorded chain and liens from the relevant county in the United States, and the registered or public price wherever the jurisdiction provides one. It costs tens of pounds or euros per title and produces a version of the asset that nobody with a position wrote.

Key the identifier into the model. The title number, kadastraal object, fastighetsbeteckning or parcel number sits on every row of the model the register can speak to: the owner, each registered lease, each charge, the last recorded price. A register fact and a model input are then the same record, and a change in one is visible against the other. This is the join the standard process never makes; the earlier note’s stamp of stratum, date and identity assumes it has been made.

Reconcile the pack to the register. Owner, encumbrances, area, price and registered lease terms are compared line by line with the offering memorandum and the rent roll, and every disagreement is listed with both sources and a named owner. The list is usually short, and it is the most useful document produced in the first week, because it is the set of facts on which the seller’s account and the independent record differ.

Brief the instruments on the disagreements. Counsel, surveyor and valuer receive the disagreement list as scope. Legal review then reads each registered covenant and easement against the intended use rather than listing it in a schedule. The surveyor is told which floor areas the building register disputes. The valuer is asked to explain the distance between the proposed price and the registered prices of the comparables, rather than to assemble comparables that close it. The instruments become tests of the disagreements instead of confirmations of the pack.

Label what the register cannot hold. Rent, most commercial lease terms, estimated rental value and every forward assumption are taken from the pack or from opinion, and the model says so on each row. The aim is to shrink the set of facts taken on the seller’s word to those for which no independent record exists, and to make that set visible to the committee.

1. 1

   Pull the register before the pack

   Title, charges, registered leases, deeds and mortgages, and the public price where the jurisdiction provides one, ordered on day one and before the data room is opened

2. 2

   Key the identifier into the model

   Title number, kadastraal object, fastighetsbeteckning or parcel number sits on every row the register can speak to, so a register fact and a model input are the same record

3. 3

   Reconcile pack to register

   Owner, encumbrance, area, price and registered lease terms compared line by line; every disagreement listed with both sources and a named owner

4. 4

   Brief the instruments on the disagreements

   Counsel, surveyor and valuer receive the disagreement list as scope, so legal review reads the covenant against the intended use and the valuer explains the gap to registered prices

5. 5

   Label what the register cannot hold

   Rent, most commercial lease terms and every forward assumption are taken from the pack or from opinion and are marked as such, so the committee sees which inputs rest on the seller's word

6. 6

   Test any AI system the same way

   If it cannot show the register line each figure was reconciled against, it has read the pack faster, not better

The four instruments stay. What changes is the order, and the fact that the register line and the model line become the same row. The disagreement list, not the pack, sets the diligence scope.

Source: Smart Bricks Research Lab

The register is the cheapest input in the process. A copy of a registered Dutch deed costs EUR 19.85 and the residential price record for a parcel EUR 3.70; official copies in England and Wales and county records in the United States are priced in the same range. The cost of the sequence lies elsewhere: in the analyst time to read the extracts before the pack, a matter of days on a single asset and of process design across a pipeline, and in the named person who owns the disagreement list and clears it. A firm that cannot say who reads the register first has not adopted the sequence, whatever its checklist says.

The registers stop well short of the model. None of them holds rent, and in England and Wales the register holds no lease of seven years or less, which for many commercial buildings means most of the tenancies. Price disclosure is residential in the open English series and in the Dutch price product, so a commercial buyer in both markets reads consideration off the deed or the title rather than off a dataset. The American record is county by county, unguaranteed, and uneven in what it discloses about consideration. Register entries can also be wrong, and a registered fact can be known and disregarded: the covenant in the 2020 Supreme Court case was read, and the developer built anyway. The sequence proposed here narrows the set of facts taken on the seller’s word. It does not make the register the truth, and a team that treats it so has exchanged one unexamined premise for another.

Nor does the sequence repair a generic model. An independent record constrains what a language model can assert about ownership, encumbrance and price; it does not improve the model’s reading of the lease, the survey or the comparables, which remain as hard as the benchmarks say. What the sequence gives a committee is a test it can apply to any system put in front of it, human or otherwise: for each material figure, show the register line it was reconciled against, and where there is none, say so. A system that cannot do this has read the pack faster than the analyst did, and has not read the asset.