Programs run experimentation as a loop, not a list.

Pick your rhythm

Run it at the cadence you pick

Pick the cadence that matches your traffic shape and your stakes. The structure is the same across all five — only the iteration period changes.

• Daily

  e.g. Reactive banner rotation

  For app categories where the audience refreshes intra-day (mobile gaming live ops, news, dating). The cadence cost is high; reserve for surfaces where the daily evidence-write outweighs the production overhead.

• Weekly

  e.g. Hero icon rotation

  The most common operator-default cadence. One week is long enough for native PPO/CSL traffic to converge to a posterior tier; short enough that a 12-week program writes 12 contributions to the per-app evidence library.

• Biweekly

  e.g. Subtitle refresh

  Used for surfaces with slower native traffic (smaller apps; niche categories). Each iteration still writes contribution; the trade-off is fewer compounding cycles per quarter.

• Monthly

  e.g. Seasonal banner

  For programs anchored to the calendar — holiday banners, seasonal feature highlights, recurring campaigns. The Program structure preserves the connection across cycles so year-over-year comparison is automatic, not reconstructed from spreadsheets.

• Quarterly

  e.g. Major positioning refresh

  For high-stakes surfaces where the cost of a wrong call is high enough that the cadence is intentionally slow. Quarterly iterations still benefit from compounding evidence — every prior quarter's program contributes to the next.

The architecture, not aspiration

How compounding works

The compounding claim is architectural, not aspirational. Three concrete mechanisms make iteration N+1 start from the evidence iteration N produced.

1. 1

   Each iteration writes contribution

   When a Program iteration finalises (PPO/CPP/SLE/CSL result lands), ASOLOOP writes the outcome as a bounded signal — traceable to the source experiment, scoped to the app and audience, revocable within 7 days. Inconclusive results write contribution too; precision-tier disclosure carries through to the next iteration.

2. 2

   The next iteration starts from accumulated signals

   When the Program triggers its next iteration, candidate generation reads from the per-app evidence library — every prior iteration's contribution, plus signals from one-off experiments outside the Program. Wider signal weight under Learning; narrower under Classic; commitment to one candidate under Agent.

3. 3

   Compounding is per-app, not global

   There is no global learning model. A Program running on App A doesn't feed candidate generation for App B. Every Program's evidence library is isolated at the data layer; cross-app pattern aggregation is a Teams capability (V2) under a separate consent surface.

The governance dial

Keep the controls at every iteration

Programs inherit every commitment in the rest of the platform: an ExperimentSchedule is a first-class, forkable entity — not a script — and every iteration’s generative output still runs through the claim-safety validator before it reaches you. The cadence is new; the discipline is not.

• Revoke a signal

  Any signal you disagree with stops influencing recommendations immediately.

  Armed7-day window

• Suspend the Program

  Pauses the next iteration; in-flight experiments finalize on schedule.

  ArmedNext cycle

• Roll back an applied winner

  Reverts a prior iteration's applied variant to the previous live listing.

  ArmedOne click

• Autopilot

  Agent-mode-only opt-in. Until you turn it on, every iteration awaits your approval.

  OffOpt-in · Agent mode

What Programs are not

• Not a global model.Programs compound evidence per-app, per-audience. There is no “industry-wide” difficulty score, no shared embedding pool, no cross-Operator pattern aggregation outside an explicit Teams consent surface.
• Not autonomous by default. Programs in Learning or Classic mode surface candidates for your review. Autopilot is an Agent-mode-only opt-in; until you turn it on, every iteration awaits your approval at the candidate-selection step.
• Not a substitute for keyword research. Programs operate downstream of keyword strategy. If you don’t have a keyword foundation yet, AppTweak / MobileAction / Sensor Tower are the right starting tools.