The number of variables you are allowed to change between one measurement and the next, if you intend to learn anything from the difference. Change two and you have a result with no cause attached to it. Change five and you have a story.
Validation method · Capital raising · Systematic tradingA founder rewrites the deck, tightens the target list, changes the subject line and switches the sending address — all in the same week. Replies double. He now believes the new deck is working. He has no evidence for that whatsoever, and the belief will cost him the next three months.
Raising capital produces very little clean data. A round is one event. You get a handful of real conversations, a smaller handful of second meetings, and a binary outcome months later. Almost nothing you do during a raise generates a sample large enough to learn from — which is precisely why founders learn the wrong things from it, confidently, and repeat them.
The failure has a specific shape and it is easy to recognise once you have seen it. Something is not working. You respond by fixing everything you can think of at once, because fixing one thing feels slow and the runway is not getting longer. The number moves. You attribute the movement to whichever change you were proudest of. Then you build the next quarter on top of that attribution.
Suppose reply rates go from 4% to 9% after a week in which you changed the deck, the target list, the subject line and the sender. Five plausible explanations exist, and only one of them is the one you believe:
The deck improved. Or the target list got narrower and you are now emailing people the offer genuinely fits. Or the subject line is doing all of it and the deck is irrelevant because nobody opened it. Or the sending domain changed and you simply stopped landing in spam. Or nothing you did mattered and the previous week was noise — 4% and 9% on a few hundred sends are not far apart in any statistical sense.
Four of those five conclusions send you in a direction that costs money. The spam one is the most brutal, because a deliverability problem masquerading as a messaging problem will have you rewriting copy for a month while the actual defect sits untouched in your DNS records.
Change one thing. Measure it against a window you did not use to form the idea. Keep it only if it survives that window. Then change the next thing.
This feels intolerably slow when you are burning runway, and that feeling is the entire reason almost nobody does it. But the arithmetic is not on the side of speed. Four changes tested one at a time take four cycles and leave you with four facts. Four changes made simultaneously take one cycle and leave you with zero facts and one belief — and you will spend the following quarter acting on the belief.
The second half of the rule matters as much as the first. You must test the idea against data you did not use to come up with it. If you notice a pattern in March and then confirm it using March, you have confirmed nothing except that you can see the pattern you already saw. This is the single most common way intelligent people fool themselves, and it does not feel like fooling yourself at the time. It feels like diligence.
This is why outreach in the Dealithic dashboard runs as a sequence rather than a send. A sequence is a structure that holds most things constant on purpose: the same offering documents, the same matched investor list, the same sending identity, across a run long enough to produce a number that means something. When you then change the opening email, the change is the only thing that moved, so the result belongs to it.
It is also why the matching engine scores an offering against a fund's stated criteria instead of ranking on how interested a fund seems. Interest is a downstream signal contaminated by everything else you did that week. Stated criteria — sector, cheque size, stage, geography — do not move when you rewrite your subject line, which is exactly what makes them useful as a fixed reference point.
We apply this rule in a second place, and it is worth naming plainly because the two companies share an owner. Obsidian Quant™ is a systematic equities trading system built by Barenberg Capital, which I also founded. It is licensed as software: a client runs it in a brokerage account they own and control, and Barenberg never takes custody of client money.
Markets produce the opposite problem to capital raising. Instead of one event and no data, you get an unlimited supply of data and an unlimited supply of patterns inside it, most of which are coincidence. The temptation reverses too. In a raise you change everything at once because you have too little information. In markets you change everything at once because you have too much, and every change looks justified by something in the history.
The rule holds in both. One change, measured against periods that were not used to form the idea, kept only if it survives. Changes are staged rather than switched on wholesale, because a system that changed in four places at once cannot tell you which of the four is responsible when the results move.
The word gets used loosely enough to be worth pinning down. The trading engine is deterministic. It does not rewrite itself, it does not learn between trades, and no model sits inside the live loop adjusting its behaviour. The recursion is in the development process around it, not in the software: the system trades, the results get analysed, an analysis produces a hypothesis, the hypothesis is tested against independent windows, and a change is deployed in stages if it survives. Then more data arrives and the loop runs again.
Frontier AI models are used heavily in that loop as a research collaborator — reading results, proposing hypotheses, arguing against conclusions. They are not the product and they are not proprietary to us. What is proprietary is the codebase, the selection and exit logic, and the validation discipline that decides what is allowed to ship. That distinction matters, because a claim that a system improves itself is a claim about the software, and this one would not survive the first serious question from anyone technical.
Write down what you expect before you change anything. A prediction made in advance can be wrong. A conclusion drawn afterwards always fits. If you did not write the number down first, you did not run a test, you ran a story.
Decide what would make you abandon the idea. Not what would confirm it — what would kill it. If nothing you can imagine would change your mind, then you are not measuring anything, and the measurement is decoration on a decision you already made.
Keep a changelog. Every change, the date, and what you expected. Six months on, this is the only thing standing between you and a confident, entirely fabricated account of why things improved. Memory reorganises itself around outcomes; a dated log does not.
None of this is complicated and none of it is new. It is ordinary experimental discipline, applied to a domain — raising money — where almost nobody applies it, because the sample sizes are small, the stakes feel too high to go slowly, and there is always a reasonable-sounding argument for changing everything at once. That argument is wrong every time, and it is wrong in a way you only discover a quarter later.
The Dealithic dashboard builds your offering documents, scores them against 50,000+ funds and family offices, and runs the outreach from your own address — sequenced, tracked, and stable enough that when something changes you know what changed. Free to start.
Start My Raise Free →Change one thing. Measure it honestly. Keep what survives.
Disclosure.Obsidian Quant™ is a product of Barenberg Capital. Dealithic and Barenberg Capital share common ownership, and this article links to Barenberg Capital's published results at quant.barenbergcapital.com/performance, where the system's methodology, results, and the disclosures that accompany them are maintained. All stock picks and target prices are generated by Barenberg Capital's proprietary AI trading system, Obsidian Quant™, and are provided for informational purposes only. This is not investment advice. Past performance does not guarantee future results.
Nothing in this article is an offer to sell or a solicitation of an offer to buy any security. Securities offerings referenced by Dealithic are made only through applicable offering documents.