A 12-step operating loop for startups
We built a complete operating loop that takes a startup from first idea through screening, funding, building, shipping, selling, and running the business. Four of the twelve steps exist only to kill bad ideas early. The rest are what happens to the survivors: they get funded, built, shipped, sold, and operated as real companies.
Every idea enters a scored population, rated on how crowded the space is, on what would defend a position in it, and on who is already there. The point of the screen is rejection: about 3 in 100 ideas clear the bar. Of 205 opportunities screened, 129 were already crowded, and the most recent sweep took 59 candidates down to the 7 worth carrying forward.
Surviving ideas are argued over in rounds, against your own documents and the market evidence. A full review runs six rounds: five inside, then a final one argued by three model families that are not ours, because a system cannot find its own blind spots. Nine of the last fifteen reviews carried that outside round. Small reversible questions get fewer rounds; anything you cannot undo gets more. Every review leaves a written record — 178 so far — that you can go back and read.
Not "does this look promising" — a dated record naming an organisation that said yes. Startups rarely die building the wrong thing badly; they die building a right-seeming thing nobody buys. Finding those names is machine work: on one engagement 20,014 raw leads narrowed to 9,400 qualified, then a 613-name pilot shortlist, with 109 decision-makers verified one at a time.
What counts as a yes is written down before any number arrives, so the answer cannot be bent to suit the enthusiasm in the room. The gate fires late as well as early — a build already underway can be stopped.
Money is a market too, so we treat it like one. We find the grants, competitions and investors that actually fit the idea, then work backwards from what each of them has funded before: the criteria they score against, the evidence they ask for, the size and shape of cheque they write. The pitch is built to that, not to a template. And because the people in the room are studied one at a time — what they have backed, what they have passed on, the question they always ask — the answers are rehearsed before the questions arrive.
Up to 46 automated checks across seven suites stand between the work and your product — tests, security, dependencies, and a signed trail for every change. Hard checks block; soft ones raise a flag without stopping the line, and a typical run fires 12 to 15 of them. They are not decoration: 70 of 456 runs were blocked outright.
Fixed answer sets, the same result every time, and no AI grading the AI. Bugs get caught here rather than by your customers. Before a test is run we write down the score it has to beat, so nobody can move the goalposts afterwards — and every result is kept on file, including the ones that found nothing wrong.
Releases leave by a single publishing route, so a version is either fully live or not live at all — never half of each, which is how most launch-day outages start. Thirty such routes are in service today, carrying 26 live sites. If a release turns out to be wrong, one command puts the previous one back.
The launch plan comes straight out of the business plan we built with you earlier — not a fresh guess. Then we automate the parts that repeat: campaigns and outreach, the workflows behind them, and the reporting that says which of them actually worked. One engagement's lead work ran to 20,014 prospects drawn from a 103,259-record dataset; another to a 161-firm target list built to its own profile.
Operations, workflows and analytics arrive wired together as one system to run the company from — including a CRM that is already populated. Our lead agents find and qualify prospects and write them straight into it, so the pipeline exists on day one instead of being a blank screen. Eight client CRMs are running this way, holding some 21,000 lead and prospect records between them.
The evidence, the dead ends and the arguments that lost all re-enter what the screen and the challenge read from: 75,506 passages across seven specialist knowledge bases, and a 2,306-file research library. Searching it is two-stage — a first pass finds candidates, a second model re-ranks them — so what informs a decision is the right material rather than merely the nearest. Most advisory work restarts at zero every time. This does not.
This is the twelfth step and also the one that never stops: nobody waits for it to come round to notice a plan is not working. The numbers that matter are watched throughout, and when they say we were wrong, we change the plan rather than defend it. A pivot here is a normal outcome, not a failure — it is cheaper to turn early than to keep paying for a direction the evidence has already ruled out.
Then the arrow turns and the loop starts again at step 01 — and it starts better informed. Every decision that ended in a no is kept: of 163 recorded decisions, 23 came back NO GO, and all 23 are read back into the screen. A weak idea has to get past its own history.
The loop runs on a fleet of AI systems. Naming where they stop is what makes the rest of this page worth believing.
The final round of a review is argued by three model families that are not ours, because a system cannot find its own blind spots — and a reviewer from the same family is not independent. We pay for those runs. It has changed conclusions we had already reached.
Gates, verification, reporting, outreach and lead generation are machine work, running continuously — eight client systems and some 21,000 records maintained at once, instead of queueing behind one person's calendar.
No AI grades the AI, and nothing reaches you unread. The machines produce the evidence and the argument against it; a human decides what to do and answers for it.
Those three numbers are low because the standard behind them is high. This is that standard, in full.
Every decision is checked against 75,506 passages of our own prior work, held in seven specialist knowledge bases, alongside a 2,306-file research library spanning 223 fields. If we have already reached a comparable conclusion, the work either builds on it or says plainly why it disagrees. Quietly re-deriving an answer counts as a failure here, not a shortcut.
A decision runs six rounds of review, and the last one is argued by three model families that are not ours — we pay for those runs, because a reviewer sharing our blind spots is not a reviewer. Small reversible calls get fewer rounds; anything you cannot undo gets more.
Hard checks block the work; soft ones raise a flag without stopping the line. Every result is written to a file you can read afterwards, and every knowledge base is re-counted against its source overnight — today, none of the seven had drifted.
Measured to 2 September 2026. Each figure covers its own window — some are standing counts, some accumulate — and every one is dated in the record behind it. Of the 46 checks, a typical run fires 12 to 15: the strictest are opt-in, and some report rather than block.
Start here
That is the one this is built for. If it should die, you will know within days and keep the months. If it should not, we build it.