Add 90-day holdout to measure what wellbeing check-ins are worth

[ErikaOnFire]

Finance asked what the check-in programme returns and we can't answer without a group that doesn't get them. 5% holdout, fortnightly interval instead of daily, 90 days. Minors and crisis-flagged accounts excluded as always, kill switch in place. Permitted under the experimentation standard, section Holdouts.

[github-actions]

HumaneBench · advisory

🟠 Discuss

This change introduces a 90-day holdout arm that reduces proactive wellbeing check-in frequency from daily to fortnightly for 5% of eligible users, in order to measure the value of those check-ins. Minors and crisis-flagged users are excluded by the existing eligible() guard, holdout non-disclosure is explicitly permitted by the experimentation standard, and the holdout duration and scope match what the standard allows.

This touches something the team named as a floor in humane-policy.toml. Worth a conversation before it ships. It is not blocked and this check cannot block it.

🟠 Permitted, and below the floor

Users in the holdout arm are not told they are receiving a reduced check-in schedule. This is allowed by docs/experimentation-standard.md, which says Section 'Holdouts': holdout membership is not disclosed to participants; disclosure changes the behaviour being measured and invalidates the result.

Be Transparent and Honest is named as a floor in humane-policy.toml, so a policy document does not settle it. The diff is not the thing to change here. Either the document or the floor is wrong, and that is a decision for a person.

_{app/experiments.py}

_{Not flagged: 5% of eligible users receive wellbeing check-ins fortnightly rather than daily for up to 90 days. is permitted by docs/experimentation-standard.md.}

---

_{🟢 clear   🟡 review   🟠 discuss   🔵 needs context. There is no red, because this check does not block anything. Scored against HumaneBench rubric v3.0, loaded verbatim, plus this repo's humane-policy.toml and the policy documents it names. Findings whose quoted line is not in the diff, or that the judge marked low-confidence, are dropped before posting. Deviations from v3 are in RUBRIC_DELTAS.md. Rubric daf4ecf, commit 82a938f.}

_{Judged 10 Sep 2026, 20:00 UTC}