---
title: 'Improvement graph'
description: 'A proposed control map showing how optimizers, evaluators, counter-metrics, auditors, and promotion gates govern changes to an AI system.'
canonical_url: 'https://darkfactory.dev/glossary/improvement-graph'
markdown_url: 'https://darkfactory.dev/glossary/improvement-graph.md'
collection: glossary
date_published: '2026-08-04T00:00:00-04:00'
date_modified: '2026-08-26T00:00:00-04:00'
---

# Improvement graph


## Definition

A proposed control map showing how optimizers, evaluators, counter-metrics, auditors, and promotion gates govern changes to an AI system. One path may optimize a target metric. Another checks regressions, cost, safety, or whether the target still represents the real objective. Edges make dependencies explicit: which evidence a node consumes, which component it may change, and which authority can promote or reverse the result.

Dark Factory Dev uses **improvement graph** as an applied extension of improvement loops and graph engineering. It is not an established industry standard or a claim that drawing more nodes creates safety. The point is to expose interactions that a single optimize-and-test loop hides.

## Why a graph can be useful

An optimizer can overfit its evaluator. Two evaluators can share the same blind spot because they use the same model, prompt, data, or environment. A counter-metric can block an apparent gain that merely moved cost or failures elsewhere. An auditor can inspect lineage after promotion. Mapping these relationships makes correlated evidence and circular authority easier to see before the system is trusted to modify itself.

## Distinguish it from nearby terms

An improvement loop changes a component, measures it, and repeats. An execution graph represents how work runs. A knowledge graph represents entities and relationships. An improvement graph is specifically about the control and evidence structure of system change.

## Check your understanding

Your prompt optimizer, judge, and regression reviewer all use the same model and examples. Does the graph contain three independent checks? No. Mark the shared dependency, freeze some evidence outside the optimization path, add a genuinely different signal where risk requires it, and name the rollback trigger.

## Related terms

- [Recursive self-improvement (RSI)](https://darkfactory.dev/glossary/recursive-self-improvement)
- [Controlled self-improvement](https://darkfactory.dev/glossary/controlled-self-improvement)
- [Graph engineering](https://darkfactory.dev/glossary/graph-engineering)
- [Independent verification](https://darkfactory.dev/glossary/independent-verification)
- [Verification gate](https://darkfactory.dev/glossary/verification-gate)
- [Reward hacking](https://darkfactory.dev/glossary/reward-hacking)

## Related factory areas

- [Feedback, learning & controlled self-improvement](https://darkfactory.dev/factory/feedback-self-improvement)
- [Verification, evaluation & quality truth](https://darkfactory.dev/factory/verification)

## Evidence and further reading

- [Turing Post: Is Graph Engineering Real?](https://www.turingpost.com/p/is-graph-engineering-real-why-everyone-is-talking-about-it)
- [Bouchard: Graph Engineering Explained](https://www.louisbouchard.ai/graph-engineering-explained/)
- [Harness Engineering for Self-Improvement](https://lilianweng.github.io/posts/2026-07-04-harness/)