Experiments

What's working

ranks on: reply (reply) · conclusive winners only

A read-only roll-up of what has actually separated. A variant is only a winner once the engine's deterministic math (frequency-of-success, Wilson interval — never an LLM) crowns a leader: every arm well-sampled AND the leader's pessimistic read clears the runner-up. Running and inconclusive experiments are counted but never shown as winners.

Outreach · reply experiments: 1 experiment · 0 conclusive · 0 still running

No conclusive winners yet — keep the experiments running. No winner is being implied.
By surface
Asset typeConclusive / experimentsAvg winning liftTrials
subject_line0 / 10

Web · conversion experiments: 1 experiment · 0 conclusive · 0 still running

No conclusive winners yet — keep the experiments running. No winner is being implied.
By surface
Asset typeConclusive / experimentsAvg winning liftImpressions
subject_line0 / 10
Intervention synthesis (Loop 1 · read-only · correlational)
Not enough data yet on any single intervention type — ranked for the human, no claim is being made.

No attributed intervention outcomes yet.

Correlational, not causal · read-only: correlational and confounded (field volatility, season, list quality); informs the human, never auto-executed

New experiment — generate variants (the model PROPOSES; calc DISPOSES the winner)

Generates a batch of variants for an asset and records them as arms (generation 0). This NEVER sends anything — a human activates & sends. The leader is then picked by deterministic math (frequency-of-success with a Wilson interval), never by an LLM, and only once there's enough data to separate the arms.

A slot only applies to a web asset type (e.g. landing_page_section, gbp_post) — it names where the winning copy renders, like home.hero_headline. The engine stores it NULL for outreach.

subject_line · exp:68245c4d38

ranks on: reply (reply) · status: draft

staging button test

Outreach experiment — ranked on the reply rate.

Not enough data yet — keep running. No winner is being implied. keep running — not enough data
ArmGenRate95% CI (Wilson)Successes / trials
var:subject_line:g0:b8dd2e1c00
[STUB subject_line variant one — Anthropic not live; fresh angle. Set ANTHROPIC_API_KEY for the live draft.]
0 · seed0.0%0.0%100.0%0 / 0
Double-down unlocks once the arms separate.