Growth Engine — Autonomous Growth Experimentation Engine

What it does

Growth Engine is an autonomous growth motor that optimizes content through controlled experiments, not gut feel. It adapts Karpathy's autoresearch loop (hypothesize → measure → learn → promote) to marketing. Each experiment opens with a hypothesis, a single independent variable, 2–10 variants, and one primary metric. After launch, data points are logged per variant; once enough samples accrue, the engine runs the statistical scoring itself.

Scoring rests on two statistical pillars: a bootstrap confidence interval (1000 resamples, no distributional assumption, yielding the 95% range of the effect) and Mann-Whitney U (a non-parametric two-sample test robust to skewed metrics). To be declared a winner, a variant must clear two gates at once: p < 0.05 (statistical significance) AND ≥ 15% lift (practical significance) — filtering out noise like "200% lift on 10 users." Statuses move running → trending → keep (winner) or discard (loser).

Winners auto-promote to a living playbook; you're expected to consult it before creating new content (apply proven rules). The engine also suggests the next highest-value experiment, generates a weekly scorecard across all channels, and raises campaign pacing alerts (on pace vs. behind target).

When to use it

• Opening and managing A/B or multivariate experiments on any channel (content, email, linkedin, seo, blog)
• Logging post-launch data points and determining the statistical winner
• Pulling proven best practices from the playbook before creating new content
• Reviewing weekly scorecards and campaign pacing health
• Do not use for: one-off content creation, non-experiment reporting, or external campaign setup

Method / frameworks

• Karpathy autoresearch loop — hypothesize/measure/learn/promote, adapted to marketing
• Bootstrap confidence interval — distribution-free 95% CI (1000 resamples)
• Mann-Whitney U — non-parametric two-sample significance test
• Two-gate winner rule — p<0.05 (statistical) + ≥15% lift (practical significance)
• Hacking Growth / ICE logic — prioritize the next experiment by impact-confidence-ease
• Learning velocity > win rate — success metric is test velocity (per 2025 growth-experimentation benchmarks)