Set the person. Watch the month
Four sliders describe who is deciding. One control sets how fast a wrong decision comes to light. The grid is thirty days of decisions: quiet cells are right, red ones wrong, amber outlines were built on a wrong one before anyone noticed. The balance at the bottom prices the difference between this person and an average one.
The person deciding
Custom · profile 6.0 / 10
Decisions a day
Every prompt, every accept or reject, every "ship it". Before AI a developer made a few dozen a day. Now the ceiling is gone.
How fast do you find out you were wrong?
+N: later decisions built on the wrong one before it is caught. Assumed root repair: 40 min; each dependent decision is then made again, at the pace it was first made.
Illustrative model · thirty days
86%chance a decision is right. 47% revisited · 25.4 person-days to repair
Thirty days · 1,800 decisions
rightwrongbuilt on a wrong one
Ten sampled decisions per day, read down each column. One column is one day. One column is two consecutive days. Totals count every decision. Same inputs always draw the same picture.
Wrong, and what it dragged along
What that costs
Illustrative model, not a measurement or forecast. One person-day means eight hours of repair effort. This is debt generated by thirty days of decisions, not hours worked by one person in that month: repayment may take a team or many later months. Decisions keep arriving while repairs are deferred; the model does not limit them to available working time. Costs are assumed: a root repair that grows with the delay, plus every dependent decision made again at the pace it was first made, counted once. It does not model decisions never made.
01 · The rule
Quality is a curve, not a switch
The four traits average to a profile. All ones decide right 20% of the time, an average profile about 77%, all tens about 99%. Multiply that across a month and small differences compound. The Pirxey preset is illustrative, not a conversion of CCAT scores into measured decision accuracy. The CCAT covers speed and reading, not your domain.
02 · The lag
Wrong decisions have children
Until a wrong decision is caught, the next ones are built on it. Same hour: one. Next sprint: fifteen. In production: forty. Cost of the fix rises with the assumed delay. These are model parameters, not observed repair times.
03 · What it doesn't say
AI is not the decider
The model writes the option. A person picks it, on the strength of what they know, how fast they think, how carefully they read and how clearly they asked. That is what the sliders weigh.
To decide well, a person has to already be four things
None of these is a prompt technique. They are who the person is on the day the question arrives — and AI made the questions arrive faster than ever.
Knowledge — wide, not just deep
The model knows the framework. The person has to know that the client's invoices run on the 25th, that the last migration broke reporting, and that 'customer' means three different tables. A decision made without that context is a guess with confidence.
Speed of thought
A factory that plans, builds and verifies in minutes gives the person minutes to decide. Slow deciders do not make worse decisions — they make fewer, and the line waits. Speed is part of quality now. We test for it: our people sit the CCAT, a cognitive aptitude test of problem solving and learning speed, and 80% of the crew score above the 90th percentile.
Reading with comprehension
The output is fluent, long and confident. The one wrong line looks exactly like the fifty right ones. Reading it properly is the whole job of the inspection gate — and it is the skill that separates 'looks fine' from 'is fine'.
Clarity of expectations
'Make it better' produces a random improvement. 'Keep the API stable, cut p95 under 200 ms, add a test for the midnight case' produces the improvement you meant. The person who can say the second sentence is worth several who can only say the first.
The compounding
Decisions stack. If each one is right 95% of the time, a chain of twenty holds together 36% of the time. At 99%, 82%. Small differences in the person become large differences in the product — which is what the ledger models as repair effort.
The lag
The cost of a wrong decision grows with how long it survives — the oldest rule in software economics. AI did not change the rule. It changed the volume flowing through it: more decisions, less time between them, more built on each one.
Why we think this is the constraint now
The ledger is a model. The pressure it models is measured — by other people, on other teams, and by us on our own product.
The real technical debt now is decision debt
AI will rewrite the code as many times as you ask. It will not go back and change what you decided, and it will not make the decisions you skipped. Every wrong or missing decision, plus everything built on it before anyone noticed, is debt, and it now builds up at the speed of AI.