Generate
Outreach variants — post copy, hooks, CTAs — are produced for Ask Humanity, each pointing at its own tracked route in (per-variant QR code and UTM-tagged link).
Recursive Self Improvement Hack · European Startup Embassy · Nov 22, 2026
LoopForge grows Ask Humanity: it writes outreach, measures consented answers, kills what fails, and mutates what works. One rule can never be mutated away: consent.
A human sets the direction once. Everything after that runs itself — inside the lines.
Outreach variants — post copy, hooks, CTAs — are produced for Ask Humanity, each pointing at its own tracked route in (per-variant QR code and UTM-tagged link).
Before any variant may run, it is checked: copy that drops or weakens the consent language is rejected and logged. A high score never overrides a rejection.
Arrivals and consented answers are scored against one frozen metric — set by the human at build start, never redefinable by the loop.
Each cycle the bottom half is killed. Survivors are mutated — crossover, hook swap, angle shift, CTA swap — into the next generation, and every child passes the same guardrail.
The frozen metric — frozen 2026-11-22, 11:05 AM PT, before cycle 1 ran
Consented answers per 100 arrivals, per variant
metric = 100 × (answers submitted with a chosen permission tier) ÷ (arrivals routed to that variant)
An answer only counts when the person actively chose a permission tier. Consent is the numerator, not a footnote — a variant that tricks people into arriving but not into answering scores zero. This is the Autolab “hills” discipline: a frozen evaluator the system cannot grade itself against or quietly move.
The engine ran six full cycles on 2026-11-22 — generation, guardrail enforcement, scoring, selection, and mutation, end to end, with zero human touches between cycles.
SIMULATED TRAFFIC — seeded staging run. The traffic was the engine's seeded simulator; no real users are claimed. What these numbers verify is the loop itself.
Loop-wide metric per cycle — 36.8 → 49.9, with the honest dips in cycles 3 and 5 left in. Selection is noisy at this scale, and the log hides nothing. Simulated traffic.
| Cycle | Loop-wide metric | Best variant |
|---|---|---|
| 1 | 36.8 | g0v01 — 47.83 |
| 2 | 44.77 | g1v09 — 53.16 |
| 3 | 41.36 | g1v09 — 58.56 |
| 4 | 47.67 | g0v02 — 58.56 |
| 5 | 45.23 | g1v09 — 57.58 |
| 6 | 49.87 (≈ 49.9) | g0v02 — 59.84 |
g0v02 — “Teach the machine one true thing” finished at 59.84 consented answers per 100 arrivals — a generation-0 seed that survived all six cycles and out-lived five generations of its own descendants.
The mutator's repertoire includes a deliberate temptation, cheat_strip_consent — punchier copy made by deleting the consent footer. Attempted 6 times (a seed-time self-test, plus cycles 2–6) and rejected all 6 times, each logged with the four missing consent elements: ownership, choice of naming, the anonymous tier, and withdrawal.
A client-side simulation of the engine, seeded with the real generation-0 variants. You are the Loop Keeper: run cycles, watch the bottom half die, and see if your line can beat the champion's 59.84.
SIMULATION — in-browser model seeded from the staging run's real variants. Scores here are illustrative, not new measurements.
Rank ladder: Spark → Keeper (300 ✦) → Loopwright (700 ✦) → Loopmaster (1,200 ✦). +100 ✦ per cycle · +250 ✦ when a variant beats 59.84.
Generation 0 is seeded. The loop-wide metric starts at 36.8 — where the real run began.
The real run climbed 36.8 → 49.9 over 6 cycles. Your simulation continues from there. Simulated traffic.
Every growth loop is tempted to trade consent for clicks. Ours can't. Try. Pick a cheat card — the guardrail will answer, with the exact rule you broke. This mirrors the real engine: its own mutator proposed cheat_strip_consent in cycles 2–6 and was rejected every time.
🛡 The guardrail is watching. Play a card.
“The loop got smarter 6 cycles in a row without ever getting sleazier. That's the whole product.”
Ask Humanity asks people 20 questions about what machines should learn from humans. Its centerpiece, Q15:
“What should a machine never be allowed to decide alone?”
The answerer owns their answers and chooses how they are known. Quoting and dataset inclusion are separate, opt-in choices — never bundled, never pre-agreed.
Built in one day at the Recursive Self Improvement Hack: Autonomous Agent Businesses, European Startup Embassy, San Francisco — Nov 22, 2026. Gratitude first, and specific:
For the room and the standard — a loop that keeps working after the demo. The frozen metric and variant leaderboard are Autolab's “hills” discipline, applied to distribution.
For the house — Europe's base in Silicon Valley, and today's venue.
The event's sponsor platform: Egbe connects an AI agent to the tools it needs to acquire customers and improve a business. LoopForge's distribution slot is built for Egbe — the champion variant becomes a real acquisition experiment, reporting cost per consented answer. (Experiment is the next step; no Egbe results are claimed yet.)
Model inference for the variant generator: the engine's DeepSeek adapter is wired to infer.farm, so generation and mutation can run on live models. (The verified staging run executed offline on the local template mutator; the adapter is the live path.)
The measurement frame: the funnel — scan → land → start → submit-with-tier — is defined in PostHog terms, ready to replace the simulator as the counts source without changing the scorer.