> For the complete documentation index, see [llms.txt](https://docs.three.dev/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.three.dev/live-scoring/failure-modes.md).

# Failure modes

three.dev groups live scoring failures into labeled failure modes, so recurring problems in production traffic read as a short list instead of individual requests.

A pass rate tells you how often production traffic fails. Failure modes tell you what keeps going wrong. three.dev groups the failures [live scoring](/live-scoring/live-scoring.md) finds into labeled **failure modes** — clusters of failures with similar defects — and keeps the groups current as new failures arrive.

## Where they appear

The **Failure modes** section on the [Overview](/exploring-your-data/exploring-your-data.md) page lists the most common failure modes for the use case, with the failure count, the share of all failures, and the change over the last seven days. Recently surfaced modes carry a **New** pill. **See all failure modes** opens the full list, most common first.

A request can carry more than one failure mode — three.dev records each distinct defect separately — so the percentages can add up to more than 100%.

## How the groups build

Failure modes need a minimum number of failures before the first grouping forms; until then the section reads *Not enough failures to group yet*. Once enough failures accumulate, three.dev builds the initial groups — *Grouping failures…* — and from then on assigns new failures to the existing modes automatically as live scoring finds them. Failures that match no recurring pattern land in an **Unknown** bucket until enough of them describe a new pattern.

Each failure mode carries a short label, a one-sentence title, and a description, all drawn from the reasoning on the failures inside it. three.dev generates the labels — there is nothing to configure.

## Reading a failure mode

Select a failure mode to open it. The detail page shows:

* The failure and request counts, and when the mode was first and last seen.
* A map that plots the mode's failures in two dimensions — click a point to open its request.
* **Tag distributions** — which slice of traffic the failure concentrates in.
* The requests inside the mode, with the score for each, sorted by how central they are to the mode.

## Next

* [Overview](/live-scoring/live-scoring.md) — the sampling and scoring that feed these groups.
* [Chat](/exploring-your-data/chat.md) — ask the assistant why a failure mode happens and continue in your AI agent.