---
title: Build your own decision rule
review-state: drafting
last-human-review: "2026-06-10"
depends-on:
  - src/pytyche/experiment/recommendation.py
owner: tradcliffe
quadrant: how-to
---

# Build your own decision rule

:::{admonition} Placeholder — not yet written
:class: note

The full recipe — implementing a custom `GraduationRule` to replace
the default `ExpectedLossRule` in sequential experiments — is coming
in a future release.
:::

This guide will cover how to implement the `GraduationRule` protocol so you
can substitute your own stopping criterion into pytyche's sequential
experiment loop. Custom rules let you encode bespoke business constraints
(budget guardrails, time-to-decision caps, asymmetric loss functions) that
the built-in `ExpectedLossRule` does not express.

## What exists today

`GraduationRule` implementations are passed to
`pt.sequential_experiment(graduation_rule=...)`; the default is
`ExpectedLossRule`. The decision-theoretic quantities a custom rule
consumes come from `posterior.recommendation_summary(...)` and the
per-segment fields on `DiscoveredSegment`.

For the decision-theoretic foundations — what each input means and when to
trust it — see the [decision-theoretic inputs](../concepts/decision-theoretic-inputs.md)
concept doc.