Start a Decision Case →
🐾3Dogs NexusStructured Decision Intelligence
3Dogs Enterprise · composite client account

Four projects, and a machine that keeps running.

By Alan Finney — Founder, 3Dogs Nexus

Not a pilot and not a demo. A working investment firm, its real pipeline, and four separate pieces of work — from a single loan tape to a valuation wave of up to eighty properties to a nationwide search that has not stopped since we switched it on.

4
projects delivered
8,304
metered model calls
40+
AI models · 3 clouds
1,312
analyst seats debated
Composite of real work Names fictional No single transaction described

⚠ How to read this page

This describes real, paid work — a composite drawn from four projects, written deliberately as one blended account. Figures are aggregated, outcomes are given as proportions rather than counts, and no single transaction, portfolio, asset, market or counterparty is described anywhere on this page.

Every name is invented — the client, the principal, the lender. Asset types and geographies are generalized. Where you see a specific number it is either a machine measurement (model calls, panels, seats, compute cost) or a ratio.

Why we wrote it this way: our own Executive Committee reviewed a more detailed draft and rejected it. Its finding was that naming things fictionally is not enough when a client's transactions are live — the shape of the deal identifies it regardless of what you call the properties. Generalizing across four projects removes that path entirely, and it happens to describe the offering more honestly anyway. Its ruling is at the bottom of this page.

The setup

The client, and the actual problem

Kestrel Harbor Partners — our invented name for a real private investment firm — buys income-producing commercial real estate and the distressed debt secured against it. Loans in default. Assets coming off a lender's books. Portfolios being cleared by receivers, servicers and special-assets desks.

Their constraint was never insight; the principals are experienced buyers who know their market cold. It was throughput against a clock. Opportunities arrive as spreadsheets, tapes and listings with hard deadlines attached, and each one needs the same disciplined work: what is the collateral actually worth, what does the payoff look like, what is the realistic exit, and what is the most we can pay and still make money. Done properly, one asset is a day's work. When the file in front of you holds eighty, the calendar wins and most of them get a gut call instead.

The second constraint was defensibility. When the principal takes a number to his investors, "the model said so" is not an answer. He needs the reasoning, the assumptions, and honest treatment of what nobody actually knows.

Four projects have come out of that so far. Each one produced something the firm still uses.

Project one

The loan tape

Several dozen defaulted and sub-performing loans on a seller's tape, each underwritten individually against an independently derived collateral value.

Project two

Nationwide search

An always-on scout encoding the firm's buy-box as inspectable code, sweeping continuously for matching assets across the country.

Project three

The valuation wave

Eighty properties valued and risk-scored in a single pass, so a large raw candidate pool could be ranked rather than skimmed.

Project four

The deep underwrite

A portfolio of operating assets, each put through a forced multi-panel debate to a decisive maximum number.

Project one

A tape tells you the debt, not the asset

A lender put a tape of loans on the market. The seller's file gives you unpaid balance, payoff, and a sentence of status — and none of that tells you the one thing that decides whether the loan is worth buying: what the property behind it is actually worth.

The method

Four models vote, independently

For each property, four separate AI models — different vendors, different clouds — valued it from live grounded research without seeing each other's answers. Their median became the working value; the spread between them became the confidence score.

Where the four disagreed sharply, that asset was flagged as uncertain rather than quietly averaged into false precision. Disagreement is information; hiding it is how you buy the wrong thing.

The discipline

Recovery value, not wishful value

An offer anchored to a hopeful market value is how buyers overpay. Each offer was anchored instead to a net recoverable value — what the collateral realistically nets after the cost, time and friction of actually getting it — then segmented by how well that recovery covered the debt.

One rule was written in and never broken: an opening offer never meets or exceeds the payoff. If the math wanted it to, that was the wrong loan.

The overlay

Location risk on every asset

Every property also went through a grounded pass on its surroundings — local demand conditions, crime environment, and the specific factors that make an asset easy or painful to resolve.

Final ranking was a product, not a score: embedded equity × certainty of that equity × ease of execution. A large theoretical discount on something nobody can value or exit ranks below a modest discount you can actually collect.

Alongside it

Two full decision cases

One deep underwrite of a single complicated credit ran 2,693 model calls across 28 models, composing 9 debate panels and 130 analyst seats — and came back genuinely split, with a substantial minority arguing to reject outright.

A second, narrower case asked how to actually negotiate with the seller: 119 calls, 10 models, 9 analysts, one approving outright and eight approving only with conditions attached. The conditions were the deliverable.

Out of that split verdict came a rule the firm now applies to everything: cap the bid at the hard floor until the dominant piece of collateral has been independently verified. The dissenting seats were not overruled — their objection became policy. Preserved dissent, promoted to a standing rule
Project two

Then we stopped waiting for deals to arrive

Underwriting a tape someone sends you is reactive. The obvious next question: could the same machinery go and find assets matching this firm's criteria, nationwide, continuously, without anyone asking it to?

Step one

The buy-box, written as code

Everything the firm knows about what it wants became an explicit, inspectable file: deal types, size band and sweet spot, preferred collateral, the signals that mean a seller is motivated, and the disqualifiers that mean don't bother.

Their own hard rules went in verbatim — institutional sellers only, and it has to be transactable from our chair, meaning anything requiring someone to physically show up is disqualified no matter how good it looks.

Step two

Grounded sweeps, around the clock

Live web-grounded searches run continuously against that thesis, nationwide, pulling candidates from wherever these assets actually surface.

Raw volume is not the product. Every candidate is scored on a deliberate blend — 45% deterministic rules, 55% model triage — with thresholds set so that a "hot" flag still means something when you see it.

Step three

Cheap detection, expensive research

Survivors get a second, costly pass: grounded research on the actual asset and market, then a re-rank by a stronger model with that new context in hand.

That ordering is what makes running this continuously affordable rather than theoretical.

Step four

A short ranked list, on a cadence

What reaches the client is not a firehose — it is a ranked shortlist with the reasoning attached, on a weekly rhythm, plus alerts when something genuinely hot appears between runs.

Categories the firm has decided to handle differently accumulate on their own watchlist instead of polluting the main list.

The shape of the funnel matters more than any single number in it. For roughly every hundred candidates the sweep surfaces:

100
surfaced by the nationwide sweep
~90
remain after removing duplicate listings
~39
are addressable at all
~21
are genuinely unique assets, not the same property listed five ways
~7
justify a full workup
~4
survive the firm's own thesis gates

That last drop is the one worth noticing. Roughly a third of fully-researched, genuinely interesting assets are killed by the client's own stated rules — an operating business rather than a passive asset, an equity interest rather than real property, a sale that requires someone to be physically present. The system's job is not to find the most deals. It is to find the deals this particular firm can actually do.

The scout has been running ever since. As of publication it holds over a thousand candidates in its pipeline, has raised several hundred alerts, has carried 188 assets through full workup, and maintains a separate 117-asset watchlist in the category the firm handles differently.

The first backtest of the scoring model failed review — it measured itself against overlap with a list we already had, which proves nothing except that a system can agree with its own inputs. The Executive Committee rejected it and required grounded enrichment plus a soak period before any accuracy claim could be made. Executive Committee ruling · our own methodology, rejected
Project three

Up to eighty properties, valued and risk-scored in a single wave

A large candidate pool is only useful if you can rank it. Skimming eighty properties tells you which ones have the best photographs.

The third project put up to eighty properties through the full independent-valuation treatment in a single wave — the same four-model blind vote per asset, the same median-and-spread confidence, the same grounded location-risk pass — so the client could rank a large pool on embedded equity, certainty and executability rather than on which listing happened to look best.

This is the piece that is genuinely impossible by hand inside a useful timeframe, and it is where the economics stop being arguable. A wave that size, each property valued four times independently and risk-scored, is roughly a month of analyst work. It ran unattended.

It also produced the thing clients underestimate until they have it: a defensible ranking. Not a list of assets, but an ordered list with the reason for each position attached, which is what actually lets a small team decide where to spend its attention.

Project four

Then the portfolio that had to be right

The fourth project was a portfolio of operating assets with a hard decision attached to each one, and the client's instruction was blunt: analyze all of them, properly, and tell me which ones the money actually works on.

"Properly" meant something specific. Each asset went through the platform's deep mode — a forced multi-pass ensemble where the system composes several independent debate panels per case rather than one, then reconciles what they disagree about. Each analysis had to cover the asset and its physical condition, the local sub-market's performance, the surrounding economy and what actually generates demand there, the immediate neighborhood, valuation on a per-unit and yield basis, a forecast of how the situation realistically plays out, a hands-off manageability score for an owner who will never live nearby, and a decisive maximum number.

They ran one at a time, sequentially, through the night — 3 to 13 debate panels per asset, composed to fit each question rather than from a template.

The panels did not agree, and that is the product

A system that returns a portfolio of confident yeses has told you nothing.

Near-unanimous rejection

14 of 15 seats: walk away

On one asset the panel was close to unanimous that the deal could not be made to work at any realistic price. One seat held out for conditions. The recommendation went out as a walk-away with that lone objection printed underneath it.

Conditional consensus

16 of 17 seats: yes, but

On another, nearly every seat supported proceeding — and every one of them attached conditions. An unconditional approval never appeared. The conditions, not the approval, were what the client needed.

Honest uncertainty

One asset returned LOW confidence

The seats split genuinely. Rather than reporting a confident-sounding average, the system capped its own confidence at LOW and said so on the first page. We build for that deliberately.

Preserved dissent

Minority positions survive to the page

Where a strong minority disagreed with the majority, its reasoning was carried into the delivered report rather than smoothed away — so the client argues with the objection instead of never seeing it.

Where the assets landed

Stated as proportions, because the specific portfolio is not the point — the distribution is.

~1 in 5

Pursue

Value after improvement clears the total cost of acquiring and fixing the asset with room to spare. Bid with confidence up to the maximum.

~1 in 3

Only at the right price

Works only below a ceiling that sits beneath what the asset will most likely sell for. Enter the disciplined number, expect to be outbid, walk the instant it runs past. Chasing them turns a maybe into a loser.

~1 in 2

Walk away

Total cost to own and improve exceeds what the asset is worth once stabilized. No bid discipline fixes that; the only winning move is to let it go.

The most valuable output was arguably the middle column. It is easy to sell a client the winners. Telling them that a third of the things they were excited about are viable only at a price they will not get — and that they should plan to lose those on purpose — is the part that saves real money, and it is the part a system optimized to please would never say.

One structural assumption was surfaced and labeled rather than buried: in this asset class the published opening figure is a teaser, and the realistic clearing price is a multiple of it. Every "all-in" number was built against the likely clearing price, not the advertised opener — and the report said so in plain language, as an assumption the client should verify rather than a fact the model had discovered.

The machinery

What all of that cost to run

Aggregated across the analytical runs in this relationship. Every figure here is a machine measurement pulled from the platform's own metering, not an estimate.

8,304
metered model calls
40+
distinct AI models
93
independent debate panels
1,312
analyst seats
MeasureAcross the engagement
Separate projects delivered4
Largest single valuation waveup to 80 properties
Independent valuations per property4 models, blind to each other
Metered model calls, analytical runs8,304
Distinct AI models drawn on40+ across three clouds
Adversarial debate panels composed93
Analyst seats debating1,312
Metered compute, largest portfolio underwrite$38.48
Nationwide searchContinuous — not included above
Forty-plus models is not a marketing number; it is what these runs actually drew on — models from more than a dozen vendors, served through Amazon, Microsoft and Google, with panel composition varying per case because the system builds the panel to fit the question.

The honest framing is not "this replaced an analyst." It didn't. A person still scoped it, still argued with it, still rejected drafts and still signed the email. What it replaced was the impossibility — the version of the month where eighty properties do not all get valued properly, so six get real work and seventy-four get a glance.

That is the actual Enterprise proposition. Not cheaper analysis. Analysis that would otherwise not have happened at all, delivered inside the window where a decision is still available to you.

The layer underneath

Connecting it to your own data

Everything above ran on grounded public research plus files the client supplied. The connector layer — wiring the analysis directly into a client's own systems — was offered as part of this engagement, and across our work generally many such connectors have been built and tested. It is worth explaining properly, because it is what Enterprise is ultimately for.

The principle is simple: your edge is your data. An analysis grounded only in what the open web knows is an analysis any competitor can reproduce. When 3Dogs reasons over your live internal data — and cites it — the output stops being generic.

Mechanically, a connector wraps one API and makes its data available to the debate as grounded, quotable evidence. Each one is private to the account, credentialed separately, encrypted at rest, read-only by default, and never co-mingled with another client's data. Adding a second source is a new isolated connector, not a rebuild.

The pattern is not industry-specific. The same mechanism applies wherever a high-stakes decision depends on data your organization already has:

IndustrySystems a connector typically wrapsThe decision it grounds
Commercial real estateListing and marketplace APIs, comparable-sales and valuation services, title and lien records, tax assessor dataAcquire, pass, or what to bid
Lending & creditCore banking, loan servicing platforms, credit bureaus, collateral and UCC filingsUnderwrite, restructure, or sell the paper
Private equity & M&AVirtual data rooms, CRM and pipeline systems, financial data providers, portfolio reportingProceed, renegotiate, or walk
Legal & investigationsE-discovery platforms, docket and filing systems, contract lifecycle management, document repositoriesCase strategy, exposure, settle or fight
InsurancePolicy administration, claims systems, catastrophe and geospatial risk modelsUnderwrite, price, reserve, or decline
Healthcare & life sciencesClinical and claims data warehouses, trial registries, supply and device telemetry (deployment tier permitting)Capital allocation, service-line and program decisions
Manufacturing & supply chainERP and MES systems, supplier-risk feeds, logistics and freight platforms, quality systemsMake-or-buy, supplier switch, capacity commitment
Energy & utilitiesAsset-management systems, market and price feeds, operational historians, outage and maintenance recordsCapital projects, asset retirement, hedging posture
Public sectorProcurement and solicitation feeds, GIS and spatial data, permit and inspection systems, open data portalsAward, prioritize, or defer
Any organizationThe data warehouse you already have — Snowflake, BigQuery, Redshift, Databricks — plus CRM and BI layersWhatever your hardest recurring decision is
This table describes capability. Connectors are offered as part of every Enterprise engagement and many have been built and tested; we are deliberately not itemizing which specific integrations exist for which clients, because that is the clients' business and the list moves. What we will not do is imply that a given integration was built for a client when it wasn't — in the relationship described on this page the connector layer was proposed and priced, and the delivered work did not require switching it on.
Your machine, your rules

Where it runs, and what runs inside it, is your decision

Two questions come up in every serious Enterprise conversation: where does this thing live, and which AI models are allowed to touch our information. Both are configuration, not negotiation.

Where it runs

Up to and including your own building

  • Fully managed — we run it; your account gets isolated storage and its own connectors.
  • Your cloud, your keys — deployed into your own cloud account, against your own model credentials. You own the infrastructure, the logs and the bill.
  • On-premises, behind your firewall — installed inside your network. Nothing crosses your perimeter, and your existing certified environment is the control boundary.
  • Sealed / air-gapped — no outbound connectivity at all, running models you host, grounded only on your own corpus. Built for classified, defense and sovereign work.
Which models run

Allow-list, deny-list, or open-weight only

The roster is not fixed and it is not ours to impose. Clients routinely constrain it:

  • Exclude a vendor outright — competitive, legal, contractual or policy reasons; no justification required.
  • Exclude a cloud or a jurisdiction — including "nothing outside our own region."
  • Open-weight only — restrict to models you can host yourself, which is what makes the air-gapped tier possible.
  • Pin an approved list — if your risk committee has already blessed a set of models, we run that set.

The method survives the constraint. Panels are composed from whatever roster you permit, adversarial roles are assigned the same way, and the debate works the same — because the process is the product, not any particular model.

What leaves

Data handling you can point at in a review

  • Zero-retention option — content held only long enough to produce your deliverable.
  • Never used for training — not ours, not anyone's.
  • Read-only by default — we pull what informs the decision; nothing is written back unless you ask for it.
  • Per-account isolation — your data is never co-mingled; no other client can trigger your connectors or see your cases.
  • Credential scoping — your keys encrypted, scoped to your account, never exposed in logs.
What you can prove afterwards

An audit trail, not a black box

  • Every model that participated in a decision, and in what seat.
  • The evidence each conclusion rests on, labeled verified, inferred, assumed or contradicted.
  • The dissent — who disagreed, how strongly, and why — preserved rather than averaged away.
  • The assumptions the analysis had to make, stated as assumptions.
  • What would change the recommendation, so a reviewer can test it rather than trust it.

This is the part that matters when someone asks six months later why you did what you did.

One consequence worth stating plainly: because the roster is configurable and the deployment can be sealed, "we cannot use external AI services" is usually a solvable constraint rather than a disqualifier. If your organization has already ruled out cloud AI, that is the conversation to have on the first call — it changes the shape of the deployment, not whether it can be done.
The part that isn't AI

The first several versions were not good enough

This never appears in a product demo, and it is the reason the tier exists.

The analysis was finished. The deliverable was not. Between the completed run and the email that finally went out, a human consultant read each draft and sent it back. These are the actual rejections:

"Why would I pay more for something worth less?"

The first draft showed, for each asset, what it would cost and what it was worth — and the "worth" figure had been derived from the cost figure. A circular comparison. It looked like analysis and contained none. Rejected outright and rebuilt around an independently derived value.

Today's value and tomorrow's value are not the same number

For an asset you intend to improve, comparing all-in cost against current value makes every value-add purchase look insane. The deliverable had to separate cost-to-own-and-fix from worth-once-fixed-and-stabilized, and show the gap explicitly. Rejected until it did.

The categories have to add up

A draft summary described recommendations whose categories did not reconcile to the number of assets analyzed. Trivially fixable, fatal to credibility. The generator now hard-fails if they don't sum.

Nobody widens a column before they judge you

A workbook column truncated its label on first open, so the most important classification in the file was unreadable until the client resized it. Sent back. Presentation is not separate from the work.

None of those were AI failures in the interesting sense — the underlying analysis was sound in every case. They were judgment failures about what makes a document trustworthy to the person receiving it, and every one was caught by a person whose job was to be accountable for what left the building.

A model can produce a defensible number. It cannot decide whether the document containing that number is honest, legible, and safe to put in front of someone's investors. That decision has an owner, and the owner is a human being. Why Enterprise includes a consultant
Governance

Our own committee reviewed this page, and cut it

What the Executive Committee ruled

The Executive Committee — four independent frontier AI models on three clouds, debating in rounds — reviews consequential decisions on this platform. It reviewed the client work described here three separate times during delivery. Its most useful ruling was to let honest arithmetic overrule the panels: where an asset's margin was negative, it was to be classified as a walk-away even if its own analysis panel had voted to proceed with conditions. Assets moved into the red column because of that rule.

Then we asked it about this page. The draft in front of it described one specific portfolio in detail, under fictional names. It rejected that draft. Its finding, after two rounds and a concession from the models that initially disagreed, was that fictional naming is not sufficient de-identification while a client's transactions are live — the shape of a deal identifies it regardless of what you call the properties.

What you are reading is the rewrite it asked for: a composite across four projects, outcomes as proportions, no single transaction described, and no figure that could be matched against a live opportunity.

There is a version of this business that publishes the impressive parts and deals with the consequences later. We had a panel of our own systems tell us that would be a breach of the thing the entire engagement runs on — trust — and we took the instruction. If that discipline is what you want applied to your own decisions, that is precisely the product.