We killed 191 of 196 ideas. The bug was in the machine, not the ideas.

The number that started this

LeanAI Studio has sourced 196 ideas since it started. 191 are dead. 5 are still alive, and every one of those 5 entered the pipeline before the window I am about to describe.

In the seven days ending August 20th, we sourced 18 new bets. We killed all 18. On August 20th alone, we sourced 7 and killed 7, most of them the same day. The deepest any idea reached across that entire week was stage 4 of a 10 stage pipeline. Nothing got anywhere near a real wedge.

Sourcing was not the problem. We were running at roughly six times our weekly target and producing nothing that lived.

Where I looked first, and where the answer actually was

My first instinct was to blame the ideas. Wrong market, wrong timing, wrong founder fit. That reading is easy, and it is also the one that lets you keep doing exactly what you were already doing.

The real audit went somewhere else: our own rulebook. On August 19th we counted every standing rule the studio runs on. 61 active rules, roughly eleven rejection rules for every one rule that generates something new. In the studio's entire history, not one of those 61 rules had ever been removed.

Each rule was correct on the day we wrote it. Every one arrived after a real mistake: a real dollar wasted, a real bad call nobody wanted to repeat. That is not the failure. The failure is that rules arrived on every incident and left on none, so the aperture kept narrowing no matter what any of us intended. Four straight weeks of zero survivors is what that narrowing looks like from the outside.

This is a rate problem, not a judgment problem. A system that adds a safeguard after every mistake and never removes one will eventually strangle its own throughput, on a timeline slow enough that nobody notices until someone counts. Every individual safeguard is defensible. That is exactly what makes the total so hard to see while it is happening.

On August 19th the active rule count dropped from 61 to 56. That is the first time it has ever gone down.

What we replaced

The old pipeline was a ten stage validation cascade: source, scope, economics, voice of customer, wedge, feasibility, regulatory, distribution, competitive intelligence, offer. An idea had to survive all ten gates to become a landing page. Nothing that week got past the fourth.

The new pipeline has four stages: MINE, DESIGN, TEST, and BUILD or SELL.

• MINE. We stopped sourcing "ideas" and started sourcing pain dossiers: 20 or more verbatim complaints about a specific, named competitor's product, counted against how many total reviews we actually looked at, dated, and checked for whether we are even allowed to act on what we found.
• DESIGN. This used to be a kill gate. Now it is a stage where we design 2 to 3 concrete propositions per qualified dossier, and every line of each one has to trace back to an actual quoted complaint.
• TEST. Ten to fourteen days, zero paid traffic, permanently. We register what counts as a pass before we launch anything. The bar is an unsolicited action: a signup, a booking, a deposit. A friendly reply to a question we asked does not count.
• BUILD/SELL. Unchanged. It was never where things broke.

What this is not

This is not a result. Engine v4 is three days old as I write this. It has not completed a single test yet. I do not know whether it works. I know the old pipeline did not, and now I know why.

The part worth stealing

If you are running anything with agents, or rules, or a team that answers every mistake with a new checklist item, ask yourself one question: when was the last time you removed a rule instead of adding one?

If the honest answer is never, you probably do not have a quality problem. You have a rate problem, and it stays invisible precisely because every individual decision that created it was right.

We are going to find out over the next few weeks whether the new pipeline produces survivors. I will write about it either way.

Sources

All figures in this post are measured directly from LeanAI Studio's internal bet ledger and a rule audit run on August 19, 2026: 196 bets sourced to date, 191 killed, 5 active; 18 bets sourced and 18 killed in the seven days ending August 20; 7 sourced and 7 killed on August 20 alone; deepest pipeline penetration that week was stage 4 of 10; the studio's standing rule count went from 61 to 56 on August 19, its first ever decrease. Engine v4 (MINE, DESIGN, TEST, BUILD/SELL) was approved August 21, 2026, and had produced zero completed tests as of this post's publish date.