Research notes

How we think when the number is not published.

Six short notes on method, each from a real engagement. They are about the decisions before the code: what to fix first, what to gate, what to derive, what to refuse to estimate, and when to stop.

Research note

Fix the Records Before You Build Anything on Them

A $2M approval, 529,213 parcels and a wedding list, built the same way: records first, then the short part

Collateral came second. Before a lender would look at a race car, it had to believe the numbers, and the team's books were accounting exports for three entities, sloppy and often just incorrect. Nothing was modeled until each entity's records were reconciled and rebuilt into statements a lender could read. The forecast sits on numbers that tie out, the debt model sits on the forecast, and a lender approved a $2M revolving facility on the cars.

Our builds ran in that order. The records were most of the work; what ran on them was the short part.

Four plain steps

Count what came in against a total someone else published. Resolve duplicates to one record. Label how sure each record is: a scenario until verified, an inference never dressed as a fact. Put a source or a name on every row. None of it waits for perfect data. The checks run inside the build, and what they cannot fix is flagged, not hidden.

ExhibitWhat was fixed before anything was built on it
CaseAs foundFixed firstThen built on it
Asset FinancingThree entities' exports, often incorrectReconciled, rebuilt into statementsForecast, debt model, a $2M approval
Farmland5,423,489 parcels from 81 county layersCounts checked against each county's total; duplicates removed529,213 parcels, three valuations, 15 tests
Pit WallFive driver lists that overlapOne set of profiles, 686 unique; wealth a scenario until verifiedThree tools, licensed to other programs
Wedding PlannersFour relatives' phones; one aunt invited twiceOne list, a link per family, a name on every actionReminders to quiet families; a sheet the family trusts
Source: the four cases named, with the figures each one publishes.

None of this is impossible by hand. It is clerical work on a scale nobody would pay for, which is why the records were in the state we found them. Written down as checks, it runs again on every rebuild, so a fix made once stays made.

What it is worth

We have no figure for what clean records saved, and we do not estimate one. The work is believed: a lender read the statements and approved. It is reused: the paddock profiles carry three tools, and the second diligence room took three days, the third one day. Its gaps are visible: the three farmland valuations agree within 20% on 34% of parcels, and the rest are flagged.

Where it breaks

It is slow, and it is never finished. The farmland counts matched the counties' own totals in the first 22 of 81 counties. Some homes were counted as farmland, a fix decided but not yet rebuilt. A clean record is a standard you keep, and the page should say where it falls short.

From the Asset Financing, Farmland, Pit Wall and Wedding Planners engagements. The same care with names is in Derive the number, where two spellings of one field were most of the work.

Research note

Scoring a List When the Thing You Need Is Not Published

Four conditions that multiply, a model that added them, and a cement company that fell from first to twenty-third

The first version of the model put a cement conglomerate at the top of a sponsor list. Billions in revenue, a family that could write any check. Zero chance they ever would, which one phone call confirmed. The model could not have known that from anything it had been given.

The job was to rank about a hundred candidates when there was budget to pursue ten. Score everyone on size and capacity, sort. The order was obviously wrong and nobody could say why. That bothered us more than the wrong answer.

Here is what was wrong. Four things had to be true at once: one person could decide alone, a budget line already existed, there was a US surface to use what was on offer, and there was a reason to bother this year. If any one is false the answer is zero. Those conditions multiply. The model added them. Adding is why a five-billion-dollar business with no marketing budget outranks a consumer brand a twentieth its size that spends seven percent of revenue on exactly this.

Gate first, then score

ExhibitThe four gates, applied before any points
GateQuestionIf it fails
AuthorityCan one person decide alone?Score capped at 45
BudgetDoes a budget line already exist?Capped at 40
ActivationIs there a US surface to use what we offer?Capped at 35
ComplianceCan they legally do this?Off the list
Then the weights: revealed behavior 35 points, budget mechanics 25, activation fit 15. Cap, do not zero: zeroing destroys the order inside the group you just deprioritized, and six months later you will want that order.

The gate definitions ship with the model, in writing, because two analysts will gate the same company differently. All of it is a spreadsheet and a short script. Change one definition, the ranking rebuilds, and it is diffed against yesterday's. No model in the loop. The judgment is in the definitions, where it can be read.

One late change: "can we reach them" had been folded into the score. That is a cost variable. It says nothing about whether they would say yes. Split out, a few candidates that had been ignored because they were hard to reach turned out to be the best on the list. The cement company is now twenty-third. The spice brand is fourth.

Where it breaks

The gates are judgment calls written down. Two analysts can still disagree on whether one person can decide alone, and the cap hides how close a capped company came.

From the Pit Wall engagement. Read the case.

Research note

Deriving the Number Nobody Publishes

Water cut and decline rate for every producing oil field, one join and two divisions away from a government CSV

The Global Oil and Gas Extraction Tracker lists every producing field on earth: country, operator, status, production. It does not list the two numbers that decide whether a mature field is worth touching for enhanced recovery, water cut and decline rate. Nobody sells them. Operators know them for their own fields and say nothing. Consultancies estimate them for a fee.

What does exist is the regulators. Norway's publishes monthly production per field going back decades, including produced water, field by field, free. Produced water over oil is water cut. This year's oil against three and five years ago is decline. The two numbers nobody publishes are one join and two divisions away from a CSV on a government website.

The screen

The tracker's field list as master. One loader per regulator, Norway first, returning water cut, decline, months of data and the source. A normalizer, because "Sleipner West" and "Sleipner Vest" are the same field. That was most of the work and all of the tedium. Then a workbook, because the people who use it live in Excel.

ExhibitWhat a cell is allowed to be
CellAllowedNot allowed
Water cutComputed from the regulator's produced-water series, with the sourceA basin average
DeclineComputed from the regulator's oil series, three and five years backA model estimate
Anything elseN/AA plausible number
All deterministic. When the workbook rebuilds, nothing is estimated. Each cell is a computed value with a source, or two letters.

Where it breaks

Only countries whose regulator publishes produced water are wired in. Right now that is Norway, with others in the queue. For every other field the water cut column says N/A, which is most cells. We thought hard about filling them with a basin average. It would make the workbook look complete. It would make the screen worse, because a plausible number in a screening tool is more dangerous than a blank. A blank sends you to go find out. A guess lets you stop.

The same rule runs through the Farmland and Diligence Rooms cases: not public is written as not public.

Research note

Same Asset, Eight Counterparties, Eight Reasons to Say No

Why the proposal stopped being a document and became a model with a cover letter

We almost sent the same document eight times. The first draft of the proposal was good: eight national oil companies, one technology, one pitch, logo swapped. What stopped it was reading the eight responses in advance. All eight were "no", each for a different reason.

ExhibitSame technology, eight different problems it had to answer
CounterpartyWhat it actually cared about
OneNot recovery rates. Keeping a field's workforce employed past the field's natural life.
TwoA board that could not approve anything without a domestic partner.
ThreeA sovereign fund that wanted dollars repatriated on a schedule.
FourBurned by a vendor in the same category two years earlier; needed to see it work on their rock.
Five to eightFour more, each its own, and four of the eight long shots from the start.
Counterparties are not named.

So the proposal became a model with a cover letter. For each counterparty, their own candidate field, screened from public data. A recovery model built on that field's reservoir, at their fiscal terms, not ours. Every number in the letter is an output of the model. No claim the spreadsheet cannot produce. A letter addressed to the person who would read it, leading with what their field does, not what the technology can do in general. Eight models, eight letters. The letters were the short part.

What we believe now: the document is never the deliverable. The thinking behind the numbers is, and the document is its cover. The reader can tell within a page whether it was built for them or for everyone.

Where it breaks

Whether eight bespoke models was right when four were long shots from the start. It cost weeks. Two tiers would have been cheaper. We would probably do it that way next time, and probably regret it on the one long shot that turned out to be real.

From a proposal engagement for an enhanced-recovery technology company. Related: Pit Wall, where every cold touch is built for one reader and one company.

Research note

Where the Data Stopped and the Law Started

Three weeks into the farmland build, the blockers stopped being technical. Each one was routing information.

ExhibitFour walls that turned out to be signposts
What stopped usWhat it was telling usWhat the code does now
Three counties host parcel data on platforms whose terms disallow automated crawlers by nameNot this way. Use the records office.Recorded as gaps, routed to a records request, never fetched
A fourth platform's terms prohibit automated agents outrightSameCounty shows grey on the map with a note saying why
Kansas publishes owner data but a statute bars using the names to market to those ownersYou may look but not callKansas is map-only; the outreach module cannot see it. A flag the code checks, not a memo
A 2022 Illinois appellate decision let an assessor refuse a GIS records requestAsk for the table, not the pictureThe request template asks for tabular assessment data and has a banned-words list
Engineering judgment, not legal advice. That is the first line of the rules file.

So there is a collection-rules file the code enforces. The fetcher refuses disallowed sources. Restricted states carry a map-only flag that downstream modules filter on. The request generator has a banned-words list. Every gap on the map has a reason attached.

AI wrote most of that code, and we would not have found the appellate decision without it. Deciding what the rules should be was ours, and slower than any of the code. We would rather build the constraint into the system than hope everyone downstream remembers a memo.

Where it breaks

This is engineering judgment, not legal advice, and terms of service change without notice. The rules file is only as current as its last review.

From the Farmland engagement. Read the case.

Research note

The Spreadsheet That Killed Our Own Idea

An insurance product for motorsport programs, stopped in a week by twenty cells and one number we could not verify

We did not build this one. That is the note.

The idea: an insurance product for motorsport programs. Track days, rally, club racing. Expensive cars, amateur drivers, scheduled events, and a gap between what a standard policy covers and what happens at a circuit. Everyone in that world has a story about the gap.

Before any code, a napkin. One tab, fewer than twenty cells. Filled in three ways: optimistic, what we actually believed, and one assumption wrong.

ExhibitThe napkin, by assumption
Assumption
Defensible?
Source
Effect if wrong
Cars per event
Yes
Stood in the paddock
Small
Premium per car per event
Yes
Public comparables
Small
Events per year
Yes
Published calendar
Small
Loss ratio
No
One proxy, carriers do not share
Believed case goes negative
Premium, cars, events, net after the carrier's share, loss ratio, return at ten percent. The one proxy for loss ratio turned the believed case negative and the optimistic case into a rounding error.

The napkin's answer: this is a business only if the loss ratio is about half what the one data point suggests, and that cannot be verified in under a quarter. We stopped. A week, not a quarter.

Where it breaks

It still nags. The gap is real. Someone with carrier data will build it, and we will have had the napkin in a folder. Fine. The napkin's job was to say whether to spend a quarter.

This was retail cover for amateur track days. Captive cover for professional team fleets is a different product, and it went ahead; see Asset Financing.

The same discipline, applied before launch, is the Stackwell case: decide what done looks like before the first dollar.