Naming what a UI change cannot move
The credibility move is subtractive: for each metric in the set, say plainly whether your change could plausibly touch it, and put the ones it cannot in writing before anyone else has to ask.
You now have five metrics defined precisely enough to argue with. This lesson does the opposite of what that setup suggests. It is about writing down, in advance and in your own document, which of them your change cannot touch.
Say plainly where this argument comes from: it is this course’s own position, not a finding from any source cited in this module. a16z, SaaS Capital, Bessemer and Feld define metrics. None of them argues for this practice. It is the synthesis of the four preceding lessons into something you can actually do on a Tuesday, and it is presented as reasoning you should be free to reject.
Why subtraction reads as strength
A proposal that claims four effects is not four times as persuasive as one that claims one. It is weaker, because a listener evaluating four claims will find the weakest one and evaluate that. The proposal is worth what its worst sentence is worth.
Worse, an unqualified claim leaves the room to fill in the scope themselves, and people fill it in generously. Say “this should help retention” and somebody hears the headline net retention number for the whole book of business. You did not say that. You will be held to it anyway, because you left the boundary undrawn and somebody else drew it.
Naming the metrics you cannot move fixes both problems in one move. It reduces the surface area of what can be attacked, and it takes the drawing of the boundary out of anyone else’s hands.
The five, run against one real change
Take the review gate: per-field confidence, a source crop, and approve-or-correct before anything downstream fires. Here is the pass, metric by metric, with the reasoning from each of this module’s lessons behind it.
| Metric | Could the review gate touch it? | What would have to be true |
|---|---|---|
| ARR | Not directly, and not this quarter | A customer who would have left renews instead, at their own renewal date |
| Gross retention | Plausibly, in one segment | Errors were a stated reason for churn among high-volume accounts |
| Net revenue retention | No claim available | The effect would be real but inseparable from expansion inside the same fraction |
| CAC payback | No | Nothing here changes what winning a customer costs, unless the gate becomes a demo asset |
| Rule of 40 | No claim available | Several links downstream through the profit term, none of them sizeable with what exists today |
One plausible claim out of five, and it is hedged by a segment and a date. That is not a weak case. It is the case as it actually is, and the version that survives someone checking.
Notice the shape of the second column. Two entries say “no,” which is a factual statement about mechanism. Two say “no claim available,” which is different and more honest: the effect might be real, and the measurement cannot isolate it. Collapsing those two into one word throws away the distinction that makes you sound like you have thought about it.
The hard case, which stays in
The advisory recommendation surface is the one that resists this exercise, and that is exactly why it belongs in your set. It recommends and does not execute. Nothing it does lands on a transaction, so the chain from the surface to any of the five metrics runs through a human deciding to act, and neither end of that chain is instrumented.
Run the pass on it and the table comes back almost empty. The correct output is a case that says: I have no financial claim for this yet, and here is the one thing I would need to observe before I did. That is a pass. A case that honestly concludes it cannot claim anything has done its job, and it is worth more to you than a case that claims something it cannot defend, because you can write the second one in five minutes and never recover from it.
There is a version of this that is not honest, and it is worth naming so you can hear it in your own drafts. “It is too early to measure” sounds like restraint and is often a way of keeping the claim alive without exposing it. Restraint says which metric, in which segment, by when. Evasion says later.
Writing the block
Four lines, in this order, sitting under whatever claim you are actually making.
- Claiming: one metric, one segment, one earliest date.
- Not claiming: the metrics from the set that this change cannot reach, and for each, one clause on why.
- Would need to see: the observation that would let a “no claim available” become a claim. This is the line that turns the block from a disclaimer into a plan.
- Would falsify: the result that would make you say the change did not work.
The last two are the ones people skip, and they are the ones that get read. Anyone can list what they are not claiming. Stating in advance what would prove you wrong is the thing that almost nobody does, and it is the strongest sentence available to you in that room.
Check your recall
Answer from memory — no scrolling back.
Retrieval check
Somebody objects: “if you keep telling us what your work does not do, why would we fund it?” Answer them.
Check your answer
The answer is about what a claim is for. You are going to check this in six months. I would rather hand you one number you can check than four you cannot, because the four cost me everything I say afterwards.
The room is not choosing between your narrow case and a bigger true one. It is choosing between your narrow case and a bigger case that will be found out. Anyone who has sat through the second version knows what it costs, and the people who have not will find out from someone else’s proposal rather than yours.
Hands on
Run the subtractive pass on all three cases
Done when: Each case in VALUE-CASES.md carries a four-line block — claiming, not claiming, would need to see, would falsify — with every one of the five metrics from this module accounted for as either claimed or explicitly declined.
- For each case, take the five metrics in turn: ARR, gross retention, net revenue retention, CAC payback, Rule of 40. Every one gets a verdict. A metric you skipped is a metric the room will assume you were quietly hoping for.
- Separate no from no claim available, and write the one clause of reasoning after each. The clause is what makes it a judgement rather than a shrug.
- Write Would need to see for each “no claim available”. One observation, stated concretely enough that someone could go and check whether it exists.
- Write Would falsify for the claim you are actually making. A number and a date. If you cannot write one, the claim is not yet a claim.
- Do the advisory surface last and let it come out nearly empty if that is the truth. Then read all three blocks aloud. The one that makes you uncomfortable to say is the one to bring into the chat first.
What this does not cover
The module ends with a set of verdicts and no document to put them in. That is the next module: one page, four fields — hypothesis, cost, mechanism, measurement — starting with why the mechanism field is the one that does the work. After that comes choosing a measurement the company can actually obtain, arriving with an answer to what you would cut instead, and the attribution trap itself, which is the lesson this whole course was written to deliver.
Read this next — primary source
State of the Cloud 2023Bessemer Venture Partners — free; a venture firm publishing the benchmarks its own portfolio is measured against
The argument in this lesson is the course’s own and does not come from this report. Read it anyway, and read it for an unusual reason: it is the tidiest available inventory of the numbers a design case gets asked to move — net revenue retention bands, payback bands, an efficiency score — laid out as though each were a settled fact about good businesses. Practising the subtractive move on somebody else’s benchmark table is the fastest way to get fluent at it. Take each band in turn and ask what would have to be true for your work to touch it. Most of the time the honest answer is nothing, and being able to say that quickly is the skill.
Stuck, curious, or think this lesson is wrong? Ask your teaching agent. The lessons are the scaffold; the conversation is where the learning gets unstuck.