Skip to the content.

Context

An LLM generates content that users consume directly (a lesson script, example sentences, a translation). Automated validation catches many errors, but some still reach users, and the users reading the output are the best-placed detectors of what slipped through.

Example

Zeeguu lets learners flag LLM-generated content as wrong, in the form that fits each surface:

Either way the report says which piece of output is wrong, not merely that something is.

Problem

How can errors that slip past automated checks be caught and pinpointed, using the users who are already reading the output?

Forces

Solution

Give users a low-friction way to flag LLM-generated output, and capture enough context to act on it. Match the report’s anchor to the shape of the content: for continuous content, anchor to the user’s current position (the point in the script); for discrete content, attach a per-item report control so the flagged unit is unambiguous.

Route each report into the quality machinery: record it against the artifact (composes with LLM Content Validation Tracking), let it trigger regeneration (composes with Soft Invalidation of LLM Artifacts), and use LLM Output Provenance to identify the prompt and model that produced the flagged output.

Consequences

Known Uses

Notes


💬 Open an issue about this pattern