---
schema: 1
type: term
title: _ground truth_
description: >-
  **_Ground truth_** is data or outcomes you already know are correct — such as
  real metrics and AB-test results — used as the reference point when evaluating
  how a prompt or model change affects outputs. When no ground truth exists, the
  channel's material shows how to construct one, for example by writing out the
  output you wished you'd received and turning it into evaluation criteria.
canonical: 'https://liquidgold.cc/channel/everyinc/explore/term/ground-truth-793e496ec2dd'
markdown: >-
  https://liquidgold.cc/channel/everyinc/explore/term/ground-truth-793e496ec2dd.md
llms: null
parent: 'https://liquidgold.cc/channel/everyinc/llms.txt'
channel: everyinc
updated: '2026-10-23T22:12:35.346Z'
language: null
children: []
---


# \_ground truth\_

Canonical: https://liquidgold.cc/channel/everyinc/explore/term/ground-truth-793e496ec2dd

Parent: [Channel index](https://liquidgold.cc/channel/everyinc/llms.txt)

## Definition

**_Ground truth_** is data or outcomes you already know are correct — such as real metrics and AB-test results — used as the reference point when evaluating how a prompt or model change affects outputs. When no ground truth exists, the channel's material shows how to construct one, for example by writing out the output you wished you'd received and turning it into evaluation criteria.

In this channel's material, **_ground truth_** is what separates guessing from evaluating. Jared Zoneraich frames the choice directly: when you backtest a new prompt against your last thousand or ten thousand responses, the signal that it _changed everything_ is not useful on its own — **with ground truth, you anchor on real metrics and AB tests** instead of eyeballing diffs or chasing similarity scores. [^d351f8cb-2125-5ece-8488-6f6ee66c9ee2@1791] The material then shows what to do when ground truth doesn't exist. For tasks like email summarization, Zoneraich recommends dissecting your own intuitive judgment into explicit heuristics — killing the “This summary is about…” preamble, fixing markdown indentation — or **handwriting the output you wished you'd gotten** and extracting principles from it. [^d351f8cb-2125-5ece-8488-6f6ee66c9ee2@2063] That handwritten ideal effectively becomes the ground truth you measure drafts against. Michael Taylor's workflow automates the same move: give the model a good and a bad version, ask what differs, and turn each difference into an evaluation criterion. [^d351f8cb-2125-5ece-8488-6f6ee66c9ee2@2063] Explore both paths and you'll see the pattern: ground truth is the anchor you either already possess (real metrics, AB tests) or must deliberately construct (heuristics, handwritten ideals, good-vs-bad comparisons) before an eval can tell you anything reliable.