CITATION.cff

cff-version: 1.2.0
title: "Cortex: a measured loop for agent context, and the lab that tests it"
message: "If you use this software or its data, please cite the paper given as preferred-citation."
type: software
authors:
  - family-names: Giampieri
    given-names: Samuele
license: MIT
version: "1.0.5"
date-released: "2026-10-05"
repository-code: "https://github.com/samugit83/cortex"
abstract: >-
  Cortex measures whether a skill, rule or instruction earns its place in an
  agent's context, by running the agent with and without it on tasks harvested
  from real sessions and comparing pass rates under fixed gates. This repository
  contains Cortex, the end-to-end lab that evaluates it, 56 scenarios across five
  task families (four house rules and a control), and the rollout-level data behind
  every number in lab/reports/REPORT.md. The gates contain no model of any kind, by design; a
  calibrated judge is confined to the proposal layer and every primary number in
  the report was produced with it switched off.
preferred-citation:
  type: article
  title: "Cortex: Measurement-Gated Evolution and Pruning of Coding-Agent Skills and Rules"
  authors:
    - family-names: Giampieri
      given-names: Samuele
  journal: "arXiv preprint"
  year: 2026
  month: 10
  doi: "10.48550/arXiv.2610.14216"
  url: "https://arxiv.org/abs/2610.14216"
  identifiers:
    - type: other
      value: "arXiv:2610.14216"
      description: "The paper's arXiv identifier"
keywords:
  - llm-agents
  - context-engineering
  - evaluation
  - ablation
  - agent-skills