loop-brief.md

Loop brief

Fill this in before starting any loop. Keep it to one page. Update it when anything changes.

Field	Your answer
Loop name
Owner (one person)
Builder (who sets it up)
Area of work
Start date / review date (use real dates; set the review date at the start, based on how long results take to arrive)

1. Goal

• The one number to improve:
• Why this number is the real goal, not a stand-in:
• Where it is measured (a source the loop cannot change):

2. Actions

• What the loop is allowed to do or change (a fixed list):
• What it must never do:
• Can each action be undone? How?

3. Checking

• How an action counts as having worked:
• What it is compared against (the unchanged version, or the period before):
• How soon results come back:
• How the result is checked (exact check, known right answer, measurement, score, reviewer, or real-world result; see THESIS §3):
• How many results are expected per week: (if judged by real-world results, aim for ~200 per version; see THESIS §6)
• Who or what checks the loop's work (not the loop itself):

4. Learning and deciding

• Where results are stored:
• How the next action is chosen from those results (for example: "suggest whatever worked most often for similar cases"):
• When to judge whether a change worked (a date or number of results, set in advance, not decided by the loop):

5. Limits

• Spend cap:
• Volume or rate cap:
• Who it may contact, and who it must not:
• Off switch (how, and who can use it):
• Record of actions (where):

6. Autonomy level

Action	Level (0–4)	Who approves

7. Stop or continue

• Success looks like (number and date):
• Stop if:
• Cost cap for the trial (AI spend and people's hours):
• Estimated cost per attempt, and per improvement that turns out to work (assume most attempts fail):

Checklist

• The goal is one number that is the real result wanted
• There will be enough results to learn from in a reasonable time (THESIS §6). If not, use AI as a helper instead, or find a faster way to check.
• Actions are small and easy to undo
• Something other than the loop checks its work
• It can be compared against the unchanged version (or at least an earlier period)
• Limits are set in the system, not only in instructions
• There is an off switch and a record of every action
• The loop uses past results to choose what to do next
• Customers, the public, money and personal data need a person's approval for each action
• No personal data goes into an AI tool without approval
• Listed and approved by someone other than the owner (THESIS §15)
• One owner is named, and it's still worth doing if most attempts fail