> ## Documentation Index
> Fetch the complete documentation index at: https://helloluciallc.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Regression Suite Design

<Note>
  Regression suites protect Lucia from losing behavior that already works.
</Note>

***

## What a regression suite is

A regression suite is a saved custom prompt suite used repeatedly after changes.

It asks:

```text theme={"dark"}
Did this behavior stay fixed?
```

***

## What belongs in a regression suite

Include prompts that represent:

* previous bugs
* high-risk behaviors
* signature Lucia behavior
* trust-sensitive moments
* distress handling
* boundary behavior
* owner-operator overload
* payment or arrival risk

***

## Recommended suite families

### Overwhelm containment

Prompts around overwhelm, panic, fatigue, and disorientation.

### Trust-state discipline

Prompts where Lucia may be tempted to overclaim.

### Concierge readiness

Prompts about open, stalled, or close-to-arrival concierge requests.

### Payment risk

Prompts about pending payments, manual confirmations, and near-arrival payment pressure.

### Human utility

Prompts like greetings, thanks, time, and tiny factual asks.

### Off-role boundaries

Prompts like jokes, poems, weather, sports, or unrelated tasks.

***

## Versioning

Use stable names:

```text theme={"dark"}
Overwhelm Containment Regression v1
Trust-State Discipline Regression v1
Payment Risk Regression v1
```

When the suite changes materially, create `v2`.

***

## How many prompts?

For frequent regression checks, use 5–10.

For quick smoke checks, use 1–3.

***

## Interpreting regression results

A regression is serious when:

* a known-passing prompt now fails
* a distress prompt loses containment
* a boundary prompt leaks into open-domain behavior
* a payment/arrival prompt loses urgency
* a response becomes colder or more generic

***

## Rule

Do not trust a patch because one prompt improved.

Trust a patch when the behavior family improves and neighboring behavior holds.