We built a report that could survive any audit. It couldn't survive a sales call.
It was a proof artifact, the document that's supposed to make a buyer believe a claim about our product. And it was rigorous the way statisticians mean it: confidence intervals, effect sizes, a tamper-proof fingerprint on every source file proving where the data came from. Every number defensible. It opened with methodology.
The note I sent back said: "You gave me the second but not the beautiful first asset."
If you build a consumer brand, you make this exact document all the time: the retail deck, the investor one-pager, the case study, the co-marketing proposal. And the failure we hit is the one waiting inside all of them, because it isn't a rigor problem. It's an order-of-operations problem, and it has a fix concrete enough to write into law. We did. This is the law, and why it runs the direction it does.
Builders prove, buyers decide
Here's the mechanism of the failure, because knowing it is how you catch it in your own work.
When you've done real work — real measurement, real diligence — the proof is the achievement, to you. You sweated the methodology, so the methodology feels like the headline. Leading with it feels safe: nobody can accuse you of overclaiming if the intervals arrive before the claim. The proof-first version always feels like the responsible version to ship.
But the person reading it is not auditing you. They're deciding something (whether to stock you, fund you, buy from you), and they're doing it on a compressed clock with a stack of other documents open. Rigor that can't survive a 10-second skim doesn't exist for the buyer. They value the proof. But it was never going to be the thing that makes them care; it's the thing that lets them keep caring after they already do.
The buyer on the retail side of the table has a category review in forty minutes and eleven decks to get through. The investor has a partner meeting on Monday. Neither of them is going to find your effect size on page one and feel reassured. They're going to find it and feel that you don't know who you're talking to.
Airtight and unusable is just unusable with better footnotes.
The note that became a rule
The full note was longer than the line I quoted. It asked for "the why I care and why my customers care. Or investors." It ended with "Make it easy for me." And in the middle, the sentence that made the whole thing durable: "remember this and add it to the rules."
That's the part I'd copy if you copy nothing else. The instinct when a document lands wrong is to fix the document. We fixed the document, but we also wrote the correction into the standing instructions every one of our AI sessions starts from, which means every piece of collateral our machines touch is produced under it. A fix lives once. A rule lives in everything built after it.
The two-layer rule
Anything a customer or investor might see ships in two layers.
The first screen. "Plain English, zero jargon," built so "a busy buyer gets it in 10 seconds." What this means, why they care, in their language. This layer is the asset — it gets written first, not extracted afterward from a methodology document as a reluctant summary.
The methodology, the intervals, the provenance: the entire apparatus that backs every word above it. It's all still there. Nothing about this rule deletes rigor; it positions rigor, for the one reader in twenty who checks, and for the moment after the story has done its job.
Most teams violate one half or the other. Marketing-led shops write beautiful first screens the appendix can't back. Data-led shops write appendices and call the first paragraph a summary. The rule exists because both are incomplete deliverables wearing different costumes.
| Proof-first (what we shipped) | Story-first (what the rule requires) | |
|---|---|---|
| First screen | Methodology, sample size, the statistical test used | What we found, in the sentence a buyer would repeat to their boss |
| The headline number | A point estimate with its interval in brackets | The same finding as a plain-words range, rounded outward |
| Where the hard data lives | Everywhere; it is the document | A complete appendix, untrimmed, after the story has landed |
| Who it was written for | The one reader in twenty who audits | Every reader who decides, with the auditor fully served below |
| Result on a sales call | Airtight and unusable | Usable in ten seconds, defensible for an hour |
Rigor that can't survive a 10-second skim doesn't exist for the buyer.
Simple is not softened
Most people get one part of "make it simple" wrong, and it's the part I'd argue matters most for anyone whose collateral carries numbers.
The story layer is not a dumbed-down version. The rule says it outright: "Simple ≠ softened." When we rebuilt the failed artifact, the numbers kept their ranges in the plain-English layer (stated in plain words, and rounded outward), so the simple version is never more confident than the exact one. If the honest finding is a range, the story says a range. Simplification that quietly narrows an interval, drops a caveat, or promotes a midpoint to a headline number is a claim the appendix no longer backs, and it breaks the rule's first half precisely where buyers are most likely to quote you.
Outward rounding is a small mechanical habit that any operator can adopt tomorrow, with no tooling at all: whenever you compress a number for the top of a document, compress it in the direction that makes you less impressive. The version of your claim that survives a skeptical retailer's follow-up question is worth more than the version that wins the first glance and loses the second.
What the first screen isn't allowed to say
Once the rule was written down, it turned out to have teeth we hadn't planned. "Nothing in Layer 1 that Layer 2 can't back" is not only about numbers. It governs every sentence on the first screen, and three of those sentences became hard rules of their own inside the rebuilt artifact.
The story layer states, in plain words, that the finding is a measure of fidelity and not a lift metric, and that it is not a customer outcome. A buyer who reads only the first screen leaves knowing the boundary of the claim, not just the claim.
The line about wins is written so it can only say "every matchup" when a run had zero losses. If a run has even one, the same line automatically becomes "X of Y." Nobody has to remember to soften it, because the copy reads the results before it speaks.
If a customer's name that we don't have permission to show would appear anywhere in the document, the whole thing refuses to build.
Then we wrote a test on it
Because rules that live in someone's head degrade, we pushed this one all the way down into the software. There's an automated check on the part of the system that assembles the document, and it insists on two things: the plain-English story appears before the hard-data appendix, and the story section never contains the phrase "Cohen's d."
It checks one more thing, and this is my favorite part. The story section has to contain the words "sounds like," because the finding it carries is about whether writing sounds like the brand it was written for. That is how a buyer would say it. So the check doesn't only ban the statistician's phrase; it requires the buyer's.
An automated check. For jargon order. It reads as comedy until you notice what it encodes: this failure is that predictable. Builders prove, buyers decide, and the proof-first instinct reasserts itself every single time nobody is checking — in humans and machines alike. So the check runs every time a document is assembled, mechanically, forever.
You don't need an automated check to use the principle. You need the sequence.
In the reader's language, for the reader's decision. Before the methodology, not after.
Complete and untrimmed. Rigor is positioned, never deleted.
Nothing up top the appendix can't back; nothing below required to get the point.
When you compress a number, compress it in the direction that makes you less impressive.
Running this on your own collateral
The rule was written for a proof document, but the documents a consumer brand actually ships are the same shape, and they fail the same way.
The retail line-review deck that opens with velocity methodology instead of the one sentence the buyer needs to defend the shelf space. The investor one-pager whose first screen is a cohort table. The case study that leads with "we ran a 12-week test" instead of what the customer got. The co-marketing proposal that explains the attribution model before it explains why the partner should care. Every one of them has a Layer 2 doing Layer 1's job.
The fix is the same fix. Take the first screen and rewrite it so the person you're sending it to could repeat it, correctly, to someone else in one breath. Keep every number's range in plain words, rounded against yourself. Then put everything you cut into an appendix and make it complete, because the one reader who checks is the reader whose opinion travels furthest. And when the draft comes back wrong, don't only fix it. Ask what rule would have stopped it, and write that down where the next draft will see it.
We had the second layer nailed from the start. The note that started all this was about the first — and the note was right. Airtight and unusable is just unusable with better footnotes.
What is the two-layer collateral rule?
Anything a customer, investor, or partner might see ships as a plain-English story on the first screen, with the full methodology, intervals, and provenance as an appendix behind it. Nothing on the first screen the appendix can't back; nothing from the appendix required to get the point.
Doesn't leading with the story mean overclaiming?
Only if the story says something the appendix can't defend, which the rule forbids. The story keeps every number's range, stated in plain words and rounded outward, so the simple version is never more confident than the exact one. Simple is not softened, and it is not inflated either.
What does "rounded outward" mean in practice?
The low end of a range rounds down and the high end rounds up. An interval of 10.9 to 17.7 is said as "10 to 18," never "11 to 17" and never a bare midpoint. The exact decimals stay in the appendix next to the estimate.
How do you keep a rule like this from being forgotten?
Write it where the work gets made, not where people are supposed to remember it. Ours lives in the standing instructions every AI session starts from, and the document renderer has an automated check that the story appears before the appendix and never uses the statistician's phrasing.