growth-loop

name	growth-loop
description	Experiment cycle for any project. Checks metrics, declares winners, proposes new experiments. Use when reviewing or managing growth experiments.
disable-model-invocation	true
context	fork

Experiment cycle. Hypothesis → test → measure → learn → repeat.

Karp Loop Integration

For bounded mutation loops (web performance, landing page optimisation, code quality), use the Karp Loop harness at ~/dev/products/elixir/karp_loop/.

Karp Loop needs 5 things to start: objective, surface, evaluator, budget, decision. If any are missing, do not start the loop — propose the missing pieces in queue.md.

To start a Karp Loop experiment:

KARP_LOOP_CONFIG=config/<experiment>.json bash ~/dev/products/elixir/karp_loop/scripts/run_loop.sh

To check status:

KARP_LOOP_CONFIG=config/<experiment>.json bash ~/dev/products/elixir/karp_loop/scripts/status.sh

Adjutant's role: decide WHAT to experiment on, write the config, monitor status. Karp Loop's role: run the iterations, verify, repair, keep/discard.

For Simple Experiments (no Karp Loop)

For A/B tests and metric checks that don't need the full harness:

Step 1: Orient

Read the project file (knowledge/projects/[project].md) and experiment files (experiments/_active.md + any per-experiment files for this project).

What experiments are running? What's their status? What metrics do we have?

Step 2: Measure

For each running experiment:

• Check the metric source (PostHog, Google Analytics, gws, or whatever the experiment file specifies)
• Record today's measurement in the experiment file
• Has it reached the decision criteria (days elapsed OR statistical significance)?

Step 3: Decide

For experiments that reached decision criteria:

• Declare a winner with evidence
• Move the experiment to experiments/_results.md with the learning
• Remove from experiments/_active.md
• Append to queue.md: "Urgent | growth-loop | [Project] experiment concluded..."
• If the winner needs to be made permanent (deploy, update config), create an Action Required item in queue.md

For experiments still running:

• Log today's data point
• If stalled >48hrs with no new data, flag in queue.md

Step 4: Learn

After processing all experiments, review experiments/_results.md for patterns:

• What's working across projects?
• What's consistently failing?
• Are there adjacent hypotheses suggested by the data?

Update the project file with any strategic insights.

Step 5: Propose

If fewer than 2 experiments are active for this project, propose a new one:

• Write a hypothesis based on learnings from past experiments
• Define the metric, baseline, target, and decision criteria
• Create a new experiment file in experiments/
• Add to experiments/_active.md
• Append to queue.md: "Action Required | growth-loop | New experiment proposed..."

The experiment is proposed, not started. Adjutant (main agent) reviews and either approves or adjusts before deployment.

Experiment File Template

# [Experiment Name]
project: [project name]
status: proposed | running | concluded
started: YYYY-MM-DD
hypothesis: [what we're testing]
variant_a: [control - what exists now]
variant_b: [treatment - what we're testing]
metric: [what we measure]
metric_source: [where to check - URL, API, command]
baseline: [current value]
target: [what success looks like]
decision_criteria: [N days of data OR >95% significance]
results:
  - YYYY-MM-DD: [value] ([note])
winner: [declared when concluded]
learning: [what this teaches us]