# AI Video Signal V23 — Experiment Lifecycle Operating Plan

## Purpose
Turn V20 optimization ideas into auditable experiments with an immutable baseline and evidence-gated Keep / Extend / Revert decisions. V23 never edits production pages automatically.

## Start an experiment
1. Load Trend Watch with a period equal to the experiment review window (21d, 30d, etc.).
2. Load a real Awin transaction export covering that same period and confirm the full window.
3. Choose a READY Optimization Queue row and click **Start & freeze baseline**.
4. V23 freezes clicks, approved/pending transactions, approved commission, EPC, CVR, trend, currency, period dates and a baseline fingerprint.
5. If the optional `EXPERIMENT_LEDGER` KV binding exists, the same record is copied to the private cloud ledger; otherwise browser localStorage + export remains available.

## During the experiment
- Apply only the one controlled change defined by the queue.
- Do not change merchant destination, traffic source, CTA wording and placement simultaneously.
- Do not alter the frozen baseline.

## Review
- Review on or within 3 days after the planned review date.
- Load Trend Watch using the same period length as the frozen baseline.
- Load the matching Awin transaction export and confirm the full window.
- Click **Evaluate**.

## Outcomes
- **KEEP**: primary metric improved and downstream economics stayed above guardrails.
- **EXTEND**: insufficient volume, pending evidence, misaligned period or inconclusive economics.
- **REVERT**: minimum evidence exists and the change materially damaged economics.
- **RUNNING**: review date not reached.

## Hard rules
- Approved commission only is realized revenue.
- Never merge currencies.
- Click growth alone cannot create a monetization win.
- Revert requires evidence; no automatic production rollback is performed.