Run a play.

Keep a play, arm it as an experiment, let the platform propose the moves, decide each one, and read the effect against the tide with the forecast's score beside it.

A play is a rule your strategy follows when a lead turns hot: when this condition holds, propose this move, forecast what it will do, and measure what it did. This guide runs one play from the moment you write it to the moment its effect is read. Every step is one call, and every answer carries its foot: what was read, the population and the control.

The map#

1. Keep the play#

Write the play once, for your brand: its goal, the condition that starts it as a predicate over your leads, the kind of move it proposes, the measure it moves, its window, its forecast rule, what it leaves alone as the comparison, and the line at which it stops. The platform keeps it as a definition with a version, and a second keep with the same content returns the first.

Call keep a play with the play. The answer is the play as kept, with its version and its readiness. A play that names a population the record does not carry is kept as not ready, with the population named, and it cannot be armed until the population exists.

2. Arm it as an experiment#

Arming makes the play live inside one strategy. Give the share of the members the play may touch and the share left alone at random. A share under one makes the play an experiment: the members left alone are the comparison the effect is read against, and nothing else is.

Call arm a play with the play, the strategy, the brand, a share of one half and a holdout of one fifth. The answer is the armed play with its share and holdout. Two things refuse an arming, each by name: a play whose goal is none of the strategy’s goals, and a play that would read the same measure over the same members in the same window as a play already armed, which the platform calls interference. Arm the second play over a different population, or place it behind the first with an edge on the window closing.

3. Read the strategy#

The strategy reads itself as a whole: its life, the game it plays, the policy it chooses by, how many plays are armed, how many moves stand, and the share it has rolled out, read from the assignments rather than typed.

Call read a strategy with the strategy’s name. The answer names the strategy’s life and its counts, and its control compares the share read from the assignments with the share the plays asked for, within the row’s tolerance or the sampling error of a fair split, whichever is wider. A control that fails is a finding, not a formatting note: the assignments do not say what the plays say.

4. The platform proposes a move#

You do not propose the first move by hand. On its beat, the platform reads every armed play, finds the leads where the condition holds, and proposes one move per match with its forecast written first: the number expected and the number that proves it wrong, each with its interval, from the like population the play’s rule names. A move with no like population is proposed with its forecast withheld and the reason.

To propose one by hand, for a lead you have in front of you, call propose a move with the play and the lead. The same match proposed twice returns the first move.

5. Decide the move#

A move waits for a person when the play’s grant is in the propose mode. Accept it and its actions run under the narrowest grant that covers them, each with its inverse captured first. Decline it and it rests declined with your reason. Stop a running move and its remaining actions never run.

Call decide a move with the move and your decision. The answer is the move in its new place, with who decided it.

6. Read the effect, and the score#

When the move’s window closes, the platform reads the effect: the change in the measure where the move ran, less the change in the measure among the members left alone, each an interval at the platform’s level. The verdict reads held when the effect clears the number that proves the forecast wrong, missed when it does not reach it, and cannot tell between. An effect whose interval lies too little above zero claims no cause, and no impact is raised from it. An effect that cannot be read is withheld, with what would settle it: more time, more members, or a cleaner comparison.

Call read an effect with the move, and call read the forecasts with the play: the answer pairs every forecast of the play with its effect, scores each by the rule fixed for its kind, and reports whether the play’s forecasts are calibrated at the level they claim. A play whose forecasts miss more often than the level allows is a play to change, not a play to run harder.

What you have at the end#

An armed play with a share and a holdout, moves proposed with their forecasts written first, your decisions on each, the effect of each against the tide, and the play’s record of forecasts against effects. Every number carries its basis and its interval, and every reading names the population it was read over.