Here’s the deal. A forecast tells you what’s likely, and I’ve never been able to do much with that. It doesn’t tell me what to change. It doesn’t tell me whether I moved anything.
The question I care about is the other one. What can I change on purpose, why should that change the odds, and how will I know, decided before the result exists?
That last part is where most of this falls apart. Mine included. Here’s one of my own, still sitting in public.
On August 2, 2026 I put a claim on my Ledger: that people who meet RosenBuds cold would find the work credible after reading the record. Then I did the part that counts. Before collecting anything, I wrote down what would prove me wrong. At least 15 cold readers across 5 audience groups, rating credibility before and after. A separate group hearing the name out loud. Survival meant at least 1.5 points of recovery on a 7 point scale with professionals, plus spoken-name confusion under 30%. Kill meant no recovery at all, or most people mishearing the name.
Numbers picked before the data existed. That is the whole idea, and it isn’t a trick.
Then the window closed on August 23, 2026. When I went looking for the data, I found none, so on September 20 I ruled that window untested and voided it rather than scoring it.
A test with no data changes nothing. I set the thresholds, and when the window closed there was nothing to score, so that claim was exactly as unproven on August 23 as it was on August 2. Writing the rules early doesn’t help if nobody collects the evidence.
So here’s the distinction I keep working on. Forecasting asks what’s likely. Probability engineering asks what I can change, deliberately, to make a different outcome more likely, with the measurement fixed before the work starts. It’s the difference between reading the forecast and deciding when to pour the concrete.
The sentence I make myself write: changing X should materially affect Y, because Z. If I can’t name the Z, I don’t have an intervention. I have a hope with a budget.
Then, before touching anything: what’s the baseline, what exactly gets watched, for how long, in which direction, and what result would make me drop the whole idea? All of it written while I still don’t know the answer. Afterward, everything looks like it worked.
Two rules I keep re-learning. Make it matter, because if the answer wouldn’t change what I do next, it’s trivia, and trivia is expensive. Make it small, because the smallest test that could still change my mind beats the big one I’ll never finish.
Sometimes the move is to wait. Waiting is a decision too. It has a price, and it should have an expiry: say what you’re waiting for, and when you stop waiting.
None of this is mine. Decision analysis has been pricing information for decades. Real options put a number on waiting. Causal inference keeps a hard line between what you saw and what you did, which is the line most business stories walk across without noticing. Furr and Ahlstrom published Nail It Then Scale It in 2011, and the order in that title is the entire argument.
What changing the odds is not: control of the outcome. I can’t promise you a result, and anybody who does is selling you something. Odds move. Outcomes still land where they land.
The hypothesis
Work designed with a stated causal claim, and with the measurement fixed before execution, produces a clearer next decision than good advice without them.
Here’s how I’ll test it. The first engagement I admit gets the card filled in at intake: the uncertain thing, the constraint I believe is real, the alternatives I could be wrong about, the causal claim, the baseline, the falsifier, and the rule for what I do with each result. Filled in before the work, dated before the work. At the end, one question. Did the next decision get clearer, and can the owner point at the evidence that made it clearer?
What would change my mind: preregistered work that produces no clearer decision than plain good advice. Or owners who can’t act on the read, which would mean the method is aimed at the wrong moment in their week.
There’s also a debt here, and I’ve started paying it. On September 20 I voided that first window, because a window with no data isn’t a result, and re-entered the same claim under the same thresholds. The new window closes October 11. Setting thresholds and not collecting the evidence is the failure I just described, and it has my name and a date on it.
I’ll report back when the first card closes, and sooner than that when the naming test’s second window closes on October 11. In the meantime, one question worth sitting with: the last real call you made, did you write down what would have proven you wrong? Before you made it?