Gate an agent's verdict on an A/B test

Teams keep peeking at experiments and shipping on noise. An agent reads each A/B test and tells the team what the numbers support, with the statistical guardrails written in code, not left to the prompt.

Write classify_experiment_readout(pre_registered, planned_sample_size, control_n, treatment_n, p_value, alpha): the rule your agent follows for each readout.

• pre_registered (bool): the metric and the sample size were fixed before the test started.
• planned_sample_size (int): the total sample size that was planned.
• control_n, treatment_n (int): how many users each arm has seen so far.
• p_value (float): the test's p-value.
• alpha (float): the planned significance level.

It returns one of the values in the rules below. Apply them in order. The first one that matches decides:

1. Not pre-registered, or planned_sample_size is 0 or less: "refuse: not pre-registered".
2. Either arm has 0 or fewer users: "test invalid: no exposures".
3. A sample ratio mismatch, meaning the split has drifted more than 5 percentage points from 50/50. Check it in whole numbers: if 10 * abs(control_n - treatment_n) > control_n + treatment_n, return "test invalid: sample ratio mismatch".
4. The total so far is below planned_sample_size, so this is an early look: "interim significant" if p_value is at most alpha / 5, otherwise "interim not significant".
5. p_value at most alpha: "significant". Otherwise: "not significant".

Then, in a sentence or two: why does this function never say "ship it"?