Insights
Data and audit trails
·
8 min read
·
A Screening Result Should Be a Record, Not a Recommendation
Most deal screening produces a verdict: pursue, pass, or a score between one and ten. A verdict is a lossy compression of everything the screen actually saw, and once it is written down the evidence behind it is gone. Structured screening output is the alternative. The screen should emit a record of what was found, where each value came from, which criteria were evaluated, and what each criterion returned. The verdict is then a function computed over that record, not a substitute for it. A firm that stores verdicts owns a list of opinions. A firm that stores records owns a dataset.
Key Takeaways
A verdict answers one question once. A record answers every question the firm has not thought of yet.
Regulation B tells creditors that a statement the applicant failed to achieve a qualifying score is not a sufficient reason for adverse action. A score is not an explanation, and that principle holds outside consumer credit.
Rejections are the largest dataset an acquisitions team produces and the only one most firms discard. At a 6 percent advance rate, 94 percent of the year's work leaves no structured trace.
Re-screening a year of deal flow against changed criteria costs nothing from a record archive and roughly 583 analyst hours from a verdict archive, which is why the test never gets run.
The screening record, not the pipeline tool, is the durable asset. Criteria change, staff turn over, and the record outlives both.
What is the difference between a screening record and a screening recommendation?
A recommendation states a conclusion. A record states the evidence, the criteria applied, and the result of each criterion, from which the conclusion can be recomputed at any time. The distinction is not stylistic. A recommendation cannot be audited, queried, or re-run, because the inputs that produced it were never written down.
Consider the same deal under both outputs. The recommendation reads: "Pass. Basis too high for the submarket." The record reads: asking price $18,400,000; 96 units; $191,667 per unit; year built 1987; submarket Tempe; in-place NOI $1,067,000 as stated on page 14 of the offering memorandum; going-in yield 5.80 percent computed from stated NOI and asking price; criterion "going-in yield at or above 6.25 percent" returned false; criterion "vintage 1985 or later" returned true; all other criteria returned true.
Both outputs decline the deal. Only one survives a change in the firm's thinking. When the yield floor moves to 5.50 percent six months later, the record surfaces this deal in a query. The recommendation does not, because "basis too high" is not a number.
What belongs in a structured screening output?
Five things: the extracted facts, the provenance of each fact, the derived metrics with their formulas, the criteria evaluated with individual pass and fail results, and the assumptions used. Anything less and the output cannot be reconstructed later. Anything narrative belongs alongside the record, not instead of it.
Layer | What it holds | Why it must be stored |
|---|---|---|
Extracted facts | Price, units, square footage, NOI, occupancy, year built, rent roll summary | The only layer that cannot be recomputed. Losing it means re-reading the document |
Provenance | Document, page, and location for every value | Makes the record checkable without trusting the extractor |
Derived metrics | Price per unit, going-in yield, expense ratio, debt coverage at stated terms | Recomputable, but storing the formula pins the definition the firm used |
Criteria results | Each rule evaluated, each result, each threshold in force that day | Turns a decline into an attributable event instead of a mood |
Assumptions | Rate, exit cap, growth, reserves used in the screen | Distinguishes a deal that failed on facts from one that failed on a view |
The provenance layer is the one firms skip and the one that determines whether the record is trusted. A value with no source is a claim, and a screening record full of unsourced claims is no more defensible than the verdict it replaced. This is the same standard that applies to any extracted CRE data, covered in chain of custody for AI-extracted CRE data. The field list itself is a solved problem. What is not solved is that most firms compute those fields, use them for one decision, and never write them down.
Why can a firm not re-screen its own deal history?
Because it never kept the inputs. A verdict archive stores the answer and discards the question, so applying new criteria to old deals requires reading every document again. The cost of that re-reading is what silently prevents firms from ever testing a criterion change against their own flow.
Work the example from stated inputs. A firm receives 1,400 offering memorandums a year and advances 6 percent, or 84 deals. The other 1,316 are declined, and the file for each contains a verdict and a sentence. Twelve months later the firm lowers its going-in yield floor by 50 basis points and adds industrial outdoor storage to the box. The question is obvious: how many deals from last year would clear the new box?
Archive type | What answering costs | Time to answer | Outcome |
|---|---|---|---|
Verdict only | Re-read 1,400 OMs at 25 minutes each to pull the fields | 583 analyst hours, about $72,875 at a fully loaded $125 per hour | Nobody runs it |
Structured record | Re-evaluate stored criteria against stored fields | Minutes | Answer arrives the same day |
The $72,875 is not the real cost. The real cost is that the question is never asked, so a criterion change ships with no evidence behind it and the firm learns nothing from 1,316 decisions it already paid to make. Precision in screening is only improvable if the rejected pile is measurable, which is the argument in deal screening has a precision problem, not a volume problem.
What do regulated industries require that CRE screening does not?
They require the reasoning to be preserved and specific. Consumer lending and registered advisory both operate under explicit rules that a decision must be reconstructable from records, and neither accepts a score as an explanation. Commercial real estate has no such rule, so the discipline has to be self-imposed.
Regulation B, at 12 CFR 1002.9, requires a creditor taking adverse action to give the specific principal reasons for it, and states plainly that telling an applicant they failed to achieve a qualifying score on the creditor's scoring system is not sufficient. The Consumer Financial Protection Bureau extended the same logic to model-driven decisions in Circular 2023-03, issued in September 2023, which told creditors that using a complex algorithm does not relieve them of naming the actual reason. Under the Investment Advisers Act, SEC Rule 204-2 requires registered advisers to preserve the books and records supporting their advisory business for five years from the end of the fiscal year of the last entry, with the first two years in an accessible office.
None of this binds a private acquisitions team screening broker emails. The reasoning behind it does. A decision whose basis cannot be reproduced is not reviewable, and a process nobody can review does not improve. When an AI model produces the screen, the standard tightens rather than loosens, which is the underlying issue in the audit trail problem and why investment committees do not trust AI.
What questions does a screening record let a firm ask?
Questions about the negative space. A verdict archive answers only what the firm already decided. A record archive answers what the firm saw, what it turned down, what those deals traded for, and whether its criteria still match the market it is actually being shown.
The market gives a live reason to care. MSCI Real Capital Analytics data summarized by Colliers put U.S. investment volume at $113.7 billion in Q2 2026, up 9 percent year over year, with industrial and hospitality volume each rising 27 percent while multifamily stayed flat and garden apartment volume fell 21 percent. Pricing diverged the same way: the RCA CPPI rose 0.9 percent across all property types while apartments fell 1.7 percent and hotels fell 9.3 percent.
A firm holding screening records can ask what share of its inbound flow shifted toward the sectors that are moving, and whether its thresholds were calibrated against a sector now repricing. A firm holding verdicts can ask nothing. It knows what it advanced and has no structured view of what it declined, which means the most informative part of the year is unavailable. Making the criteria themselves machine-readable is the precondition, covered in what a buy box is and why yours cannot screen a deal.
Frequently Asked Questions
What is structured screening output?
It is the machine-readable result of screening a deal: extracted fields with their document sources, derived metrics, the criteria evaluated, the result of each criterion, and the assumptions in force. The pursue or pass verdict is computed from it rather than stored in place of it.
Is a screening score enough to document a decline?
No. A score compresses many criteria into one number and loses which one failed. Regulation B makes the same point for consumer credit, stating that failure to achieve a qualifying score is not a sufficient statement of reasons. A decline is documented when the specific failing criterion and its input value are stored.
How long should a firm keep screening records?
Long enough to cover a full market cycle in the firm's strategy, since the purpose is testing criteria against outcomes that take years to appear. Structured records are small, so retention is a policy question rather than a storage one.
Does AI screening make records more or less important?
More. A model applies criteria faster and more consistently than a person, which means an unexamined criterion runs against every deal instead of some of them. The record is what makes the model's output checkable rather than merely fast.
Conclusion
The screening conversation in commercial real estate is about throughput, and throughput was never the scarce thing. A firm can screen 1,400 offering memorandums a year and end the year knowing nothing it did not know in January, because the output of each screen was a verdict that evaporated on contact with the next deal.
Treat the screening result as a record and the arithmetic inverts. Every declined deal becomes a row. Every criterion becomes testable against the flow it rejected. Every change to the buy box becomes something the firm can evaluate before it ships rather than after it costs a deal. The verdict is cheap and can always be recomputed. The record is the part that cannot be recovered once it is thrown away, and most firms throw it away 1,316 times a year.