The daily testing pipeline for constructs

Testing openers by hand never works for long. Somebody has to read the numbers, decide what won and remember to start the next test. Nobody does, and the team sends the same note for a year.

Stage 1 runs that loop itself, once a day, on rules you can read.

What the pipeline does

Every first message in Stage 1 is drafted from a messaging construct, a recipe for the note. Constructs sit in a pipeline. Proposed, then queued, then testing, then champion. Or retired, or paused.

Each door runs its own pipeline, with one champion, up to two constructs in testing and the rest waiting in the queue. A new champion in one door never retires another door's.

Once a day, the pipeline looks at the record for every platform construct and moves what the evidence says to move.

The rules

The record is first messages, counted by person. Success is a positive reply or a booked meeting. Sends from example workspaces never count.

Safety comes first. Once a live construct has 50 sends, Stage 1 compares its opt-out rate to the usual rate across all first messages. At double the usual rate, it's paused, whatever its reply numbers say. A construct that annoys buyers doesn't get to keep testing.

Then the tests. A construct needs 100 sends before its record can move it. From there, Stage 1 works out how likely it is that the construct's true success rate is above the champion's. At 95 percent or better, it becomes the champion and the old champion is retired. At 5 percent or worse, it's the one retired. If it reaches 300 sends with no clear answer either way, it's retired too. A test that can't decide is taking slots from one that can.

Then the queue. Any open testing slot goes to the next construct in line, until two are testing.

And a floor. If a door ever ends up with no champion, its original built-in construct takes the seat back. For cold first messages, that's the Hypothesis Opening.

You can see every move

Every change is written to Insights, under What Stage 1 changed, with the reason. A line might read like this.

Their words back. Retired. No clear result after 300 sends.

The constructs board under Insights, Constructs shows each construct's record and how likely it is to beat the champion.

What it doesn't touch

A firm's private constructs, the ones written in the Construct Studio or proposed from your own messages, learn only from your sends. The daily pipeline leaves them alone. You turn them on, make one your default or retire it.

Nothing new enters the platform queue without approval either. A variation Stage 1 proposes waits until someone approves it.

The owner of a workspace can turn testing off. Every first message then goes to the champion.

Why the rules are strict

Small numbers lie. A few meetings from a few dozen sends looks like a winner and usually isn't. The 100-send floor and the 95 percent bar are there so a lucky week can't crown a new champion.

The opt-out guard is there for the other reason. In a services sale, your market is a short list of senior people. Burn them and they're gone for years. A construct that wins on replies and loses on opt-outs is a bad trade, so it stops. More on what opt-outs tell you in what unsubscribe requests teach you.

The pipeline runs every day whether anyone looks or not. The best recipe keeps the seat until something better proves it.