Dashboard / Optimizer & growth memo

Optimizer & growth memo

Results is where you find out what the system has been doing. Five tabs, and they are ordered from what happened, through what it tried, to whether any of it can be believed.

Analytics

The funnel itself: first open through completed, activated, trial and paid, per flow and per branch. Described in Analytics.

A/B tests

Experiments you or the optimizer started: arms, allocation, current readings and the decision metric each one is being judged on.

Optimizer

The loop's working surface — opportunities found, hypotheses formed, patches validated, and the autonomy grant each flow is running under.

Growth memo

The weekly written account. What it did, why, what it learned, and what it is not entitled to claim.

Measurement

The holdout and the reconciliation panel: claimed cumulative lift next to holdout-measured lift.

The loop, in the order it runs

  1. Detect. Find where a flow is losing people, and rank the openings by what is worth attacking.
  2. Hypothesize. Propose a specific change with a reason, drawing on the pattern library rather than starting blank.
  3. Validate. Check the patch against your optimization policy and locks. An illegal patch is not rejected at review time — it cannot be expressed.
  4. Plan. Work out what can actually be measured at your traffic in 28 days, and pick the decision metric accordingly. See Evidence & claims.
  5. Run. Open the experiment, allocate traffic adaptively, watch the guardrails.
  6. Judge and promote. Decide, and — at L3 — promote the winner.
  7. Learn. Write the outcome to the pattern library, so the next hypothesis on a similar screen starts better informed.

The growth memo

Once a week, in writing. It is the surface the whole product is designed around: if the system is going to act without asking, the minimum it owes you is a readable account of what it did.

A memo names:

  • Every autonomous action, and the reason recorded with it.
  • Experiments opened, promoted and rolled back.
  • What it learned, and where a reading was too weak to conclude anything.
  • Anything that stopped it — a guardrail breach, a budget ceiling, a kill switch — and why.

Rewrites, and how they enter

When a flow stops producing winners — several experiments in a row with nothing to promote — the optimizer stops patching it and concludes the design itself is wrong. It writes a fresh candidate from scratch.

That candidate does not replace the old flow. It enters as an experiment arm against it and has to win. The retired flow keeps its history — a new flow gets a new id, because a screen id is an analytics key and renaming one would silently corrupt every past read — and the memo links the two as successive generations of the same job.

Turning it down

Autonomy is per flow with an account ceiling, and both are editable here. Lowering the ceiling takes effect immediately. It does not interrupt anything already being served — stopping the optimizer never stops your flows from rendering.