The strongest objection first
Here is the case against this whole page, stated as an acquisitions lead would state it, because they have earned the right.
Your model of intake is backwards. Shannon proved that feedback does not increase the capacity of a discrete memoryless channel. A long enough code over a fixed corpus of comparables achieves the same asymptotic accuracy as any adaptive scheme. What you are calling a structural advantage of continuous intake is really just an engineering convenience — smaller models, shorter lag, nothing that the mathematics distinguishes as fundamentally different. Underwriting has run for decades on quarterly comp sets, trailing rent rolls, and static cap-rate tables, and it has worked. The loop you want to close was never open in the way you claim it was.
This is not a weak objection. It is Shannon's actual theorem, correctly cited, and it should make the reader expect the rest of this page to lose the argument. It doesn't, but the reason it doesn't is narrower than the thesis wants it to be, and that narrowing is the honest part of the work.
Where the objection is right
The no-increase result is real. For a channel that is memoryless — each use independent, statistics fixed and known — feedback buys nothing in the limit. A sufficiently long block code, decoded once, hits the same capacity as any scheme with acknowledgement. If real estate markets behaved like a memoryless channel — each quarter's data an independent draw from a fixed distribution of demand, supply, and rate — then a big enough comparable set decoded once would be exactly as reliable as a permanently open feed. The objection would win outright.
It also lands on the redundancy point that isn't in the quoted challenge but sits underneath it: comp sets are not thin. A given submarket generates listing data, tax records, MLS history, broker chatter, and prior appraisals, all pointing at similar numbers. That is real redundancy, and it does real work. An acquisitions lead who cross-checks five sources before underwriting a deal is running a genuine code, not superstition.
Where it breaks
Real estate is not memoryless, and its statistics are not fixed. Permit filings today shift the distribution of supply eighteen to thirty-six months out, in a way that today's comparable sales cannot yet reflect, because nothing has been built or leased yet against which to compare. Migration data shifts the distribution of demand with a lag of its own — households relocate before rents move, and rents move before cap rates re-rate. The channel has memory: this quarter's noise is correlated with next quarter's, because the same permitting office, the same builders, and the same underlying labour market generate both. Shannon's no-feedback result is stated for exactly the case that does not hold here.
Worse, the redundancy that survives the first objection fails the second one, because the sources are not independent. Comparable sales in a submarket are appraised by firms that consult the same three data vendors. Broker opinions of value anchor on the last closed transaction, which anchored on the one before it. Ten sources agreeing on a cap rate is not ten independent channel uses; it is one signal, copied. Majority decoding over correlated copies converges on the shared error, not the truth, and it does so with the same confidence a genuinely independent sample would produce. This is the mechanism behind the failure that acquisitions leads recognise on sight: the valuation model holds steady through a demand shift that permit filings had already announced months earlier, because every comp in the set was quietly downstream of the same stale assumption.
The instrument the objection has no answer for
An acquisitions lead who has been burned by this failure does not fix it by gathering more comps. They fix it by re-querying the channel: pulling current permit filings against the position they underwrote six months ago, checking whether migration data has moved since the last rent roll, and testing the rate curve assumption against this week's Treasury print rather than the one baked into the original pro forma. That is feedback in Shannon's exact sense — the receiver's prior estimate determines what gets checked next, and the check changes the estimate. Schalkwijk and Kailath's result on feedback channels is the quantitative version of what this buys: error probability can fall doubly exponentially in the number of rounds of query, which matters enormously when the deal has to close in a quarter, not in the asymptotic limit where block-length arguments live. Capacity was never the binding constraint on an acquisitions desk. Delay is. A ten-year DCF that is asymptotically correct and six months late to notice a supply shock is worthless at the point where the wire has to move.
The three generations, stated once
A Large Language Model, applied to underwriting, is a single-block decoder: a corpus of market reports and historical comps, encoded once, with no channel back to the permit office or the rate desk. Whatever was stale at the training cutoff is frozen into every valuation it produces. A Large World Model closes the loop for the duration of a site visit or a live diligence window — it can re-query the rent roll, walk the comparable, test an assumption against a document in front of it — but the return path shuts when diligence ends and the file is closed. The third position on this axis, a Large Universe Model, is the configuration in which the return path never closes: listing flow, permit filings, rate curves, and migration data stay live streams rather than snapshots, each claim in the valuation carrying provenance back to the specific filing or print that produced it, so that when the permit data and the comp set disagree, the disagreement can be traced to which channel is stale rather than averaged away.
| intake structure | failure mode | |
|---|---|---|
| Large Language Model | one block, decoded once | frozen comp set never learns the permit surge happened |
| Large World Model | loop closed for the deal window | re-queries during diligence, goes silent after close |
| Large Universe Model | loop never closes | disagreement between comp set and permit stream stays traceable, continuously |
The objection that survives
There is a third challenge worth taking as seriously as the first two: a receiver that chooses what to re-query will tend to re-query what confirms its existing position. An acquisitions lead convinced a submarket is undervalued can keep pulling comps and permit data selectively until the feed agrees with them. That is not improved reliability. It is a locked loop, and it is a real failure mode of continuous intake that a frozen corpus, whatever its other faults, does not have in the same form.
Two things follow from taking this seriously rather than dismissing it. Provenance stops being a nicety and becomes the load-bearing part of the design — every figure in the model needs a traceable path back to the filing, print, or listing that produced it, precisely so a locked loop can be diagnosed after the fact by someone other than the person who locked it. And exploration has to be engineered rather than assumed: a re-query policy that only checks the comps favouring the existing thesis is a design failure, not a neutral consequence of having a feed at all. Forced sampling — a standing rule to pull the permit data for adjacent submarkets even when the desk isn't looking to buy there — is the closed-loop answer to a closed-loop pathology. It has no equivalent in the frozen-corpus case, because a corpus cannot be diagnosed for a bias it has no mechanism to notice, let alone correct.
The claim that actually holds
The result is narrower than "continuous intake is always better," and that is the honest place to land. Shannon's theorem does not say feedback beats redundancy in general; for a known, memoryless channel it says the opposite. What it says, applied here, is that real estate's observation channel — permits, rates, migration, listings — has memory and drifting statistics, and for exactly that channel, redundancy without feedback converges confidently on correlated error while feedback converges, more slowly perhaps but correctably, on the truth. On the intake axis specifically, there is no fourth relation a valuation model can have to its market beyond decode-once, decode-while-present, and decode-continuously-with-provenance. The third is not a better version of the first two. It is the point past which the only remaining improvements are quantitative — faster feeds, better provenance, smarter re-query policies — because there is no further category of loop left to close.