The revenue model decides

The bench in session 3 scored revenue fit against a recurring-membership model, and that choice was shown on screen rather than hidden inside the arithmetic. It is worth being blunt about what follows from it: every single revenue-fit score on that bench would move at a business with a different model. Not some of them. All of them are readings taken against one sentence, and changing the sentence changes the readings.

That is not a weakness in the bench, it is the whole reason the second gate exists. The data gate in session 4 is close to objective: either the business holds the entity in full or it does not, and two reasonable people looking at the same field will usually agree. The revenue gate is a judgement, and it is a judgement about the business rather than about the use case. Which means it is the one gate you cannot outsource, cannot copy from another company's roadmap, and cannot buy from a vendor who does not know how your money arrives.

The practical consequence is small and annoying: you have to write the sentence first. Not the mission, not the strategy deck, the sentence. "We are paid a monthly fee and we keep about 70 percent of members past their first year." "We are paid once, per job, and about one customer in twenty ever comes back." Those two sentences produce different roadmaps out of an identical candidate list, and everybody in the room can check a fit score against them in about four seconds.

Paid a recurring fee for continued access. Everything that protects the next renewal is near the top: churn scores, health scores, reasons for cancelling, onboarding friction. Anything that raises the size of a single purchase is a long way down, because a bigger basket does not extend the relationship.

Paid once per transaction, with no repeat relationship. Exactly reversed. Basket size, conversion on the page, pricing at the moment of sale, fraud at the moment of sale. A churn model has nothing to hold on to, which does not stop people asking for one, and that is the subject of the fourth build-along pass.

Paid on volume you do not control, such as usage or claims. The money moves with throughput and with loss, so forecasting demand and detecting the expensive tail matter most. Retention often matters less than it feels like it should, because the customer is locked in by something other than satisfaction.

Paid on an asset being available. Utilities, rentals, fleets, equipment. Availability is revenue, so failure prediction and maintenance scheduling go to the top of a list where they would be mid-table almost anywhere else, and the same equipment-sensor case that is an expensive curiosity at a gym chain is the core of the business here.

Write yours in one of those four shapes if you can. If your business genuinely has two, score twice and keep both columns, because a candidate that scores high on one and zero on the other is a real strategic argument you have just made visible rather than a scoring problem.

Erik Brynjolfsson, Lorin Hitt and Shinkyu Yang, "Intangible Assets: Computers and Organizational Capital", Brookings Papers on Economic Activity, 2002. It works across 1,216 firms from 1987 to 1997, with a matched survey subsample of 272 used to measure organisational practice.

The headline association: each dollar of computer capital was associated with roughly $10 to $12 of market value, against about $1 for other physical capital. On its own that number gets quoted as a reason to spend on technology, which is a misreading, because the paper's substantive finding is the condition attached to it. The premium was strongly conditional on complementary organisational investment: decentralised decision rights, team structures, training. Firms with the spend and without the organisational work did not show the same premium.

So the honest way to use it is as an argument about complements. The gate check in session 4 and the fit score tonight are both, in that light, part of the organisational work rather than a delay before the real project. Nobody builds a decentralisation programme out of a business case, but the habit of deciding what a case is worth before building it is the same family of thing.

The mismatch. That paper is about organisational complements to IT investment. The sequencing rule is about revenue models deciding which case pays. Those are different claims about different objects. The paper does not measure revenue models, does not rank use cases, and could not support a statement about sequencing even in principle. Attaching it to the rule as a citation would be a false footnote, and false footnotes are worse than none because they survive being challenged for a while.

The temptation to attach it anyway is strong and worth examining, because it is the same temptation that produces the failure-rate numbers in part 3. A rule that sounds right feels safer with a paper next to it. A reader skims the citation and reads the rule as established. Nobody lied, and the page is now making a claim it cannot defend.

The test that catches it. Ask what the source would have had to measure for the claim to follow. For the sequencing rule, that would be something like: firms sorted by revenue model, comparable use cases deployed at each, returns compared. If no source in your hand did that, no source in your hand supports the rule, however adjacent the subject. Say "this is our own rule" and move on. It costs one sentence and buys you the standing to challenge everybody else's citations.

Shankar, Garcia, Hellerstein and Parameswaran, "Operationalizing Machine Learning: An Interview Study", 2022. 18 machine learning engineers, semi-structured interviews, IRB-reviewed, grounded-theory coding over 1,766 coded segments. It is a qualitative study and it says so, which is precisely why it is useful here: it is careful about what it can and cannot claim.

On the 90 percent figure it is unusually direct, and this is the sentence worth carrying around: "various articles claim that it is a problem for 90% of models to never make it to production... but we find that this statistic is misguided". The reasoning matters more than the verdict. In real practice, constant experimentation is expected to produce many model versions that never ship. That is what experimenting is. So a high non-deployment rate is not itself evidence of failure, and treating it as one gets the causality backwards: the teams doing the most careful work will have the worst-looking ratio.

On time to production, the study gives you one thing and only one thing: an attributed anecdote. Participant P19 reported that "around 40 or 50% will make it to initial launch", and that staged deployment "took over three months". That is one engineer describing one setting. The paper does not claim it generalises, and neither does this page, so if you repeat it, repeat the attribution with it.

The trace. Find the earliest thing that states the number rather than repeating it. Follow the citation chain backwards, one hop at a time, and read what each hop actually says. You are looking for the point where the chain stops, and the interesting part is what is sitting there when it does: a survey with a method, or a sentence somebody said out loud.

Rot pattern one: the expiry nobody noticed. The number was a forecast with a stated horizon. The horizon passed. Nobody reread the original, so it now circulates as a description of the present. Tell: the original wording contains a date or a phrase like "through" or "by", and the requoters have dropped it.

Rot pattern two: the remark that grew a method. Somebody says a figure on a panel or in an interview. An article reports it faithfully. The next citation drops the "said on a panel" and keeps the number. Five hops later it is "research shows". Tell: nobody can name the sample size.

Rot pattern three: the misread denominator. The number is real but counts something other than what it is used for. Non-deployment counted as failure is the clean example: the denominator is every experiment, and experiments are supposed to not ship.

Rot pattern four: the round number. 90 percent, 80 percent, two thirds. Genuine measurements are rarely round, and a suspiciously round figure with no confidence interval anywhere near it has usually been rounded by a requoter rather than measured by anybody.

None of this is about being pedantic in meetings. It is about what happens when one person in the room has done the trace and you have not. Your whole case gets re-examined on the strength of the weakest number in it, and the weakest number was the one you added for colour.

The failure rate was doing a job on that slide: it was establishing that this is hard and that a method is therefore warranted. You can do that job with your own material, and your own material is far harder to argue with because it is about the business in the room.

Put the blocked count on it. From session 4: how many of your candidates cannot be built today with the data you hold. That is a measured fact about your business, produced in an afternoon, and it makes the case for a method more forcefully than any industry percentage.

Put the zero-fit favourite on it. The most-requested item on the list, next to its revenue fit. If that fit is low, you have demonstrated the need for a second gate using the room's own enthusiasm as the evidence.

Put the two definitions on it. If two teams define your most-used metric differently, that screenshot is the strongest slide in the deck and it cost you ten minutes.

All three share a property the borrowed statistic does not have: nobody can dispute them by looking something up, because they are about you. The worst anybody can do is disagree with your judgement, and that conversation is the one you actually wanted.