> ## Documentation Index
> Fetch the complete documentation index at: https://docs.writerzroom.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Experiments
> Test headlines, CTAs, structures, voice, and channel versions — and promote a pattern only when the numbers support it.
**Experiments**, under Intelligence in the sidebar, records a test of one dimension of a piece — a headline, a CTA, a structure, a voice, or a channel-specific version — and decides whether the result supports acting on it.
WriterzRoom does not split traffic. Exposures and conversions are entered or imported from the channel that served them; what the Experiment Lab owns is the decision, not the delivery.
## The verdict
Each experiment needs exactly one control and at least one challenger. The evaluation runs a pooled two-proportion z-test and returns one of four verdicts:
| Verdict | Meaning |
| ------------------- | --------------------------------------------------------- |
| `winner` | A challenger beat the control at the corrected threshold |
| `inconclusive` | The test ran and nothing cleared the bar |
| `insufficient_data` | The test could not validly run at all |
| `no_control` | The experiment lacks exactly one control and a challenger |
`inconclusive` and `insufficient_data` are kept separate on purpose: the first means keep collecting, the second means the sample cannot support the question yet.
### Three guards
* **Normal-approximation validity.** Rather than an arbitrary impressions floor, the test requires at least 5 expected conversions and 5 expected non-conversions per arm under the pooled rate. Below that no p-value is reported, because one would not mean what it claims to.
* **Multiple comparisons.** Testing several challengers against one control inflates the false-positive rate, so the threshold is Bonferroni-corrected by the number of challengers. Adding a fifth headline makes winning harder, not easier.
* **Direction.** The p-value is two-sided, so a variant that performs significantly *worse* is never reported as a winner.
Relative lift is reported as undefined against a control that converted nobody — a move from 0% to 2% is an unbounded ratio, not an infinite improvement.
## Promotion
A concluded experiment freezes its verdict, whatever it is. An inconclusive result is worth keeping: it is the record that stops the same hypothesis being re-run and read hopefully next quarter.
Only an experiment that established a winner can be promoted into a pattern. Promotion is refused outright otherwise, with the verdict and its reason — a warning would be something to click through, and the result would be a coin flip becoming a house style rule nobody can trace.
A promoted pattern carries its evidence forward — sample size and p-value — and can be scoped to verticals, channels, or audiences. Retiring a pattern deactivates it rather than deleting it, so work already written under it stays traceable.
Scope a promoted pattern to the audience it was proven on.
Group an experiment's pieces under the initiative that ran it.