The doctrine optimises for years one through five. The bill arrives in year six. By year ten the platform team is paying for decisions it did not make, with interest.
The original argument.
Move fast, break things, learn from production. In a startup with no users and no revenue, this is rational: the cost of a broken thing is small, the cost of slow learning is existential. The argument is a conditional. It depends on the system being small, the user base being forgiving, and the engineering team being the same team that wrote the original code.
What changes around year five.
Three things, simultaneously. Revenue crosses a threshold where the cost of a broken thing exceeds the value of the next learning. The original engineers leave; the new ones inherit code without the context that justified its shape. The user base loses tolerance: the same outage that was charming at month six is a contractual failure at year five.
What we have inherited.
- A payment integration with no idempotency. The original team handled retries by reading their own production database for "did I already send this?": a pattern that worked at 100 transactions per day and did not survive 100 per minute.
- A configuration surface that lives entirely in environment variables, with no schema, no validation, no audit trail. Adding a new flag requires a deployment to read it. Renaming one breaks production silently.
- A "temporary" SQL view from 2017 that is now load-bearing for two reports the finance team runs at month-end. Nobody knows who wrote it. The git blame points at a person who left in 2018.
- A test suite that runs in 47 minutes and passes 91% of the time. The 9% is "flaky." Flaky is "we cannot tell whether the system is broken." Flaky is the worst possible state.
The repair cost.
The repair is never a single project. It is a multi-quarter discipline of paying down the most operationally-expensive item every sprint. We have measured this in production; the cost-of-repair runs at 30–40% of platform-team capacity for 18–24 months before the platform stops actively bleeding. The remaining 60% is keeping the lights on during the repair.
The lesson, stated negatively.
We refuse to treat the original doctrine as a universal. It applies to a specific stage and a specific scale. Past that stage, the same behaviour produces compounding harm, and the team paying for the harm is rarely the team that earned the early speed.
Speed in year one is borrowed against discipline in year six. The interest rate is steep.