A play is a named strategy move Deloosh can propose when the evidence fits; it uses approved moves instead of inventing new ones. Guardrails are the governed conditions and constraints attached to that move.
Plays come from a closed catalog
The strategy engine selects from an authored catalog. Its play families cover moves such as sunset, re-engagement, engaged expansion, deliverability defense, intent risk, and variant insight.
Each Signal names the recommended play and carries catalog provenance. If the active catalog is unavailable, Deloosh labels the fallback basis rather than presenting it as the active governed version.
You cannot create arbitrary plays in Deloosh v1. That closed-world rule keeps the engine from improvising a strategy move simply because it sounds plausible.
Guardrails define when a play fits
A governed rule tests the known condition against the play’s requirements. The Workbench can show the play’s family, basis, governance, guardrails, teaching, and evidence pointers so you can inspect why it was proposed.
Some plays also depend on data that is not connected yet. In that case Deloosh can lock the play and identify the missing source instead of guessing. See SIG-10, Why is a play locked (“connect data”)?.
Tenant control is tuning-only: you can adjust supported thresholds, but you do not author new plays. Even when a play fits, it remains a recommendation until you decide; see SIG-1, What is a Signal? and TRUST-1, What Deloosh does on its own — and what it always asks first.