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

# Quality gates

> How individual rule scores aggregate into an accept or reject verdict, and the arithmetic behind each gate.

The quality gate turns a set of rule scores into one verdict. It is **deterministic arithmetic** — no model is involved at this step.

## The three gates

| Gate       | Accepts when                                                       | Character                                          |
| ---------- | ------------------------------------------------------------------ | -------------------------------------------------- |
| `strict`   | Every score is 4 or higher                                         | One miss sends the item to a person.               |
| `balanced` | At least 70% of scores are 4 or higher **and** no score is below 2 | The default. Any outright failure vetoes the item. |
| `lenient`  | The average score is 3 or higher                                   | A failure can be outvoted by enough passes.        |

If no rules apply, the gate passes — there is nothing to fail.

## The arithmetic worth knowing

<AccordionGroup>
  <Accordion title="Under strict and balanced, any single failure rejects the item" icon="triangle-exclamation">
    A score of 1 on any pass/fail or deterministic rule trips the floor. Deterministic rules are therefore hard gates under these two settings — which is usually what you want, since they only fail on things that are objectively wrong.
  </Accordion>

  <Accordion title="Under balanced, middling scores drag the ratio" icon="calculator">
    Custom-scored rules returning 2 or 3 do not trip the floor, but they do count against the 70% threshold.

    With only three rules, a single custom rule scoring 3 already means just 67% are at 4 or above — a reject. Few rules combined with custom scoring makes for a twitchy gate. Know the arithmetic before mixing them.
  </Accordion>

  <Accordion title="Under lenient, a hard failure can pass" icon="circle-exclamation">
    Scores of 5, 5 and 1 average 3.67, which is accepted **despite** the failure.

    Never use `lenient` alongside brand-critical deterministic rules.
  </Accordion>
</AccordionGroup>

## Choosing one

`balanced` is the default and right for most configurations.

Move to `strict` when every rule is genuinely blocking. Use `lenient` only when the rules are advisory.

## Calibrating to your economics

<Info>
  What happens to a rejected item decides the right strictness.

  In Speedtrain's normal flow a rejection lands in a human review queue, which makes false rejections cheap and **missed violations** the expensive error. That argues for keeping deterministic rules strict and AI rules narrow.

  If rejection instead triggers costly retries, bias the other way.
</Info>

## Routing

The gate's verdict drives what happens next. Accepted items can auto-approve; rejected items go to human review with the failing rule's feedback attached, where a person can accept, decline, or retry.

All decisions happen in the Speedtrain UI. [Review sessions](/speedtrain/key-concepts/review-sessions)