Justification with a closing date
Evidentialism holds that whether a belief is justified depends entirely on the evidence held at the moment of belief, nothing else. Richard Feldman and Earl Conee stated the thesis in 1985: the justification of a doxastic attitude toward a proposition is fixed by the evidence possessed at that time. The consequence that matters for property investment is temporal. Evidence once held and since superseded does no justificatory work now. A valuation is not judged by whether it was once well supported. It is judged by whether the evidence supporting it is still current, and if it is not, the valuation has already stopped being justified — whether or not anyone has noticed.
Real estate underwriting streams four kinds of evidence continuously: listing flow, permit filings, rate curves and migration data. Each has its own decay clock. A comp set goes stale in weeks. A rate lock goes stale the moment the ten-year Treasury moves. Migration data, drawn from IRS county-to-county files or postal change-of-address records, lags by quarters but moves markets by years. Permit filings sit in between: public, dated, searchable, and often the earliest hard signal that supply is coming before a single unit leases up. This is the evidence class where the argument bites hardest, because it is the one most often filed and forgotten rather than filed and missed.
The failure, specified
An acquisitions lead underwrites a multifamily asset in a growth submarket using a rent-growth assumption of four per cent, drawn from trailing twelve-month comps and a migration read that is nine months old at the time of modelling. The model holds through underwriting, through committee, through close. What the model does not hold is the county's permit register, which recorded filings for 3,400 competing units across three sites eight months before the deal closed. Those permits were public record the whole time. They were not incorporated. Eighteen months after close, absorption in the submarket has collapsed against the new supply, effective rents are flat against a modelled four per cent, and the asset is marked down. The valuation model held through a demand shift that was visible in permit filings months earlier. Nothing about the model's original logic was wrong. Its evidential base simply stopped matching the world it was pricing, and nothing in the model's architecture was built to notice.
Two defensible readings of the same failure
Here the case splits, and it splits along a genuine fault line inside evidentialism itself, not a manufactured one.
The first position says the acquisitions lead was justified at the time of the decision. The underwriting used an institutional-grade comp set, the most recent migration release available, a locked rate curve, and no piece of evidence then in hand contradicted the rent-growth assumption. Evidentialism asks only what was possessed at the moment of belief. By that standard the belief was properly formed. What happened afterwards is not a failure of justification; it is simply the world moving, which every synchronic theory of justification permits. Blaming the lead for not having the future is a category error.
The second position says the lead was never justified, because the permit filings were not future evidence — they existed, dated and public, before the deal closed. Evidentialism distinguishes evidence one holds from evidence one merely could have held, but "possession" is doing real work in that distinction and it is not obvious it should be read narrowly. If evidence sitting in a searchable county register counts as available to a diligent underwriter, then failing to retrieve it is not bad luck striking after the fact. It is a gap in the evidential state at the moment the belief was formed, no different in kind from ignoring a comp that was sitting in the same file drawer as the ones that were used.
The permits were on the county website. Nobody has to imagine them into existence after the fact. They were there to be pulled, the same week the comps were pulled.
That objection deserves to stand unanswered for a moment, because it is right about the facts and the disagreement is not about facts. It is about what "possessed" means inside the theory: evidence physically retrieved and reasoned over, or evidence reasonably retrievable at the time. Feldman and Conee's original formulation does not settle this cleanly. Their target was reliabilism, not the internal geography of an underwriting desk, and the thesis was built to say that justification is not about external processes, not to resolve how wide the boundary of "held" evidence should be drawn inside a single decision.
Why this doesn't resolve as cheaply as it looks
The temptation is to pick a side and declare the acquisitions lead either exonerated or negligent. Resist it. Both readings survive contact with the case, and the disagreement between them is a live dispute inside evidentialist theory about the individuation of evidence, not a defect in the theory that a real estate example happens to expose.
But the two readings converge on the same practical remedy, and that convergence is the point worth extracting. If justification depends narrowly on what was pulled into the file, the fix is to make the permit feed part of what gets pulled automatically, every cycle, without depending on an analyst's initiative. If justification depends more broadly on what was reasonably available, the fix is identical: build the evidential architecture so that "available" and "possessed" collapse into the same thing, because the feed is always ingested and nothing is ever left sitting in a register waiting to be searched. Whichever reading of evidentialism a reader prefers, the remedy for this class of failure is the same architecture. That is a narrower claim than "evidentialism proves continuous intake is required." It is the claim that continuous intake is the remedy under either live reading of what possession means, which is most of the way to the same conclusion without pretending the internal dispute is settled.
Where the lineage sits under this
A Large Language Model trained on a corpus with a cutoff reasons from listing flow, permit history and migration data as they stood on that date, indefinitely afterwards. It cannot distinguish, from inside its own weights, a rent-growth assumption that is still supported from one whose support expired the month a permit register updated. A Large World Model can sense a current scene — a live feed of a specific submarket's filings, say — but only while the scene is open; close the session and the evidential state does not persist into next quarter's committee memo. A Large Universe Model is the configuration in which permit filings, rate curves, migration releases and listing flow are ingested continuously, retained with provenance — which office filed it, which county register, which vintage of the rate curve — and revised when superseded.
| evidential state | failure mode in this domain | |
|---|---|---|
| Large Language Model | frozen at a cutoff | rent-growth assumption never updates once trained |
| Large World Model | current only while sensing | permit spike seen live, forgotten at session close |
| Large Universe Model | continuous, provenanced, revised | supply shift enters the model the week it's filed |
This is not a claim that the third configuration is smarter. It is a claim that it is the only one of the three whose evidential state can, in principle, match the county register in real time rather than on a modelling cycle.
The two objections worth taking seriously here
A reliabilist would say none of this depends on evidentialism at all — what matters is whether the underwriting process is a reliable belief-forming mechanism, not whether the analyst can cite the evidence. That is a fair challenge, and reliabilism is a live alternative, not a strawman. But a process tuned on a submarket's 2022 absorption pattern is reliable only relative to that submarket's 2022 conditions; applied unchanged to a submarket that has just absorbed 3,400 new units, it is not reliable, it is formerly reliable. Reliabilism arrives at a currency requirement by a different route, but it arrives there.
The stronger objection is about cost and noise. Ingesting every permit filing, every rate tick, every migration release across every submarket a portfolio touches is expensive, and unfiltered intake imports contradiction — duplicate filings, withdrawn permits, revised county estimates. A tightly curated quarterly data pull, audited by a human analyst, may produce a better-justified valuation than a torrent of unreconciled feeds. This is correct, and it is the real constraint on any system built along this axis. The answer is not that more data beats curation; it is that provenance has to be part of the specification rather than an afterthought, so that a permit filing later withdrawn is down-weighted rather than either ignored or treated as gospel. That keeps the cost real and the deployment hard. It does not reopen a fourth evidential category above continuous, provenanced intake — it prices the third one honestly.