Cohorts + experiments

Behavioral cohort scope

When editing a bet's predicted impact, an Amplitude cohort scope picker lets you choose one of your project's Behavioral Cohorts. Picking a cohort tells Findry: "when computing variance for this bet, only count events from users in this cohort."

Common use cases:

• The hypothesis targets new users (cohort: signed up in last 30 days) — scope the bet so the predicted impact only measures against that slice.
• The bet is a mobile-only change — scope to the mobile users cohort to avoid the desktop baseline diluting the signal.
• The bet is a power-user feature — scope to the high-engagement cohort to measure impact on the audience that will actually notice.

How cohort scope works

The cohort picker in the predicted-impact editor calls GET /api/3/cohorts via the workspace's Amplitude connection and presents the list of behavioral cohorts. The selected cohort ID is stored with the bet.

When an outcome-sweep measurement run fires for a bet with a cohort ID, the Amplitude adapter appends the cohort filter to the segmentation query as the s (segments) parameter:

[{"prop":"amp_cohort_<cohortId>","op":"is","values":[1]}]

This is equivalent to checking "filtered by cohort" in Amplitude's segmentation UI. Only events from users currently in the cohort are counted.

If the cohort changes

Amplitude Behavioral Cohorts are dynamic — a cohort defined as "users active in last 30 days" has different membership tomorrow than today. The cohort ID on the bet stays fixed; the cohort filter at measurement time uses whatever membership Amplitude reports at that moment. This is intentional: you want measurement to reflect who actually experienced the change, not who was in the cohort when the bet was written.

Link an Amplitude Experiment

When creating or editing a Test, set the type to A/B experiment and the tool to Amplitude. An experiment picker appears and loads the active and recently-stopped experiments from your Amplitude Experiment project.

Linking an experiment gives you:

• The experiment's current state (running / stopped / retired) synced nightly into the Test's status.
• A direct link from the Test card to the experiment in Amplitude — one click to the rollout configuration.
• An audit trail: amplitude.experiment_synced events show every state transition in the workspace audit log.

Management API key

Amplitude Experiment uses a separate API — the Experiment Management API — which requires its own key. The main API Key + Secret Key pair can't access experiments; you need to add a Management API key.

1. In Amplitude: Experiment → Settings → API Keys → Create API Key
2. Choose Management key type. Copy the key — shown once.
3. In Findry: Settings → Integrations → Amplitude tile → Management API key. Paste and save.

The Management API key is stored encrypted separately from the main credentials. Audit action amplitude.management_key_saved is written on each save.

Nightly sync

Once a Test is linked to an Amplitude Experiment, a cron at ~06:50 UTC pulls each linked experiment's current state and updates the Test.

State mapping:

• Amplitude running → Test status active
• Amplitude stopped → Test status completed
• Amplitude retired → Test status archived

If the experiment is not found (deleted in Amplitude), the sync writes an amplitude.experiment_sync_failed audit event and leaves the Test status unchanged. Findry never auto-deletes a Test.

What syncs

The nightly cron syncs state only. It does not pull results, variant exposure counts, or statistical significance from Amplitude Experiment — those live in the experiment platform and are best viewed there.

Outcome measurement (variant lift, significance, verdict) continues to go through the main Amplitude analytics API + your metric mappings, not through Amplitude Experiment's results API. The two systems complement each other:

• Amplitude Experiment — rollout configuration, targeting rules, early-stop guardrails, per-variant exposure counts.
• Findry + Amplitude analytics — canonical metric definition, baseline vs post-ship measurement, business-level verdict, audit trail.