Run the loop

Run the Loop

Propose, pre-register, run, decide, log. Monotony is the feature. The loop was written so an agent can operate it — and so nobody, agent or human, can quietly cheat it.

Gate

Requires gate: aha, gate: keystone, and gate: instrumented all open in GROWTH.md. gate: loop is the derived summary of those three: if all three are open and loop still reads closed (or is missing), set gate: loop = open yourself and proceed; if any is closed, name it and route there. Missing gate lines are closed — never treat an absent gate as permission. Two standing sub-gates apply inside the loop:

• Virality gate: unless GROWTH.md explicitly says gate: virality = open, any proposal whose mechanism is get-users-to-get-users goes to ### Parked: virality — not debated, not refined, parked (absent or malformed line = closed). The gate opens only after activation and retention experiments run green two consecutive cycles.
• Alienation test: every experiment answers, pre-launch: would this alienate the user if they fully understood what it does? Yes or unsure → killed or escalated to a human. Non-delegable to agents.

Inputs

• GROWTH.md (keystone, ledgers, gates), experiment scaffolding from instrument-the-truth, and candidate sources: lore-ledger unknowns, activation friction, context findings, team proposals.

Steps

1. Propose. Gather candidates; rank by expected keystone impact ÷ effort. Apply the two sub-gates now — park virality mechanics, kill alienators, and log both actions (a refused experiment is a trust-ledger credit).

2. Pre-register, then launch — in that order. Per experiment, commit to the ledger before any flag flips: hypothesis, metric, decision rule, sample/duration, expected trust debit/credit. Git history is the ordering proof — a pre-registration that postdates its launch is void.

3. Run without peeking. Evaluate exactly when and how the pre-registration says. Early stopping only if the pre-registration defined it.

4. Decide and log. keep / kill by the committed rule, no relitigating. Append: result, decision, learning, actual trust debit/credit. Killed experiments get their learning written down — they are the loop's tuition.

5. Tend the gates. After each cycle: update green-cycle counts; open gate: virality when earned (and say so — it's a milestone); check the keystone's sunset date and gaming-vector tripwires, routing to find-the-keystone if either fires.

Conventions

• Rank by keystone impact, not by interestingness. A boring ledger of small honest results is what winning looks like.
• One experiment per hypothesis. Bundles produce unattributable results.
• Agents may operate every step except the alienation verdict and consent changes; every entry carries agent: or human: attribution.

Edge Cases

• "Just this one growth hack." The parked list is the answer, verbatim, plus when the gate opens. Banning the conversation, not just the tactic, is the design.
• Results ambiguous under the committed rule. Log kill (the default under ambiguity), note what a sharper pre-registration would have asked, and re-run if the hypothesis still ranks.
• Someone wants to overturn a kill with new analysis. New analysis = new pre-registered experiment. The ledger never re-tries the old one's verdict.
• Trust ledger runs net-negative two cycles straight. Stop proposing; audit with invalidate-the-lore — the candidates are coming from somewhere rotten.