CITATION.cff

cff-version: 1.2.0
title: >-
  surface-bench: a pre-registered, provider-agnostic benchmark of
  documentation drift in LLM coding agents
message: >-
  If you use this dataset or benchmark, please cite it using the
  metadata below.
type: dataset
authors:
  - family-names: McDonald
    given-names: Connor
repository-code: "https://github.com/Connorrmcd6/surface-bench"
url: "https://github.com/Connorrmcd6/surface-bench"
abstract: >-
  A controlled, single-variable benchmark measuring how documentation
  accuracy changes an LLM coding agent's task performance, and through
  what mechanism. Documentation accuracy is the only manipulated variable
  across five matched conditions (code only; code + stale doc; code + fresh
  doc; code + stale doc + an automated drift report; code + stale doc + a
  generic warning), with fully deterministic grading and no LLM judge. The
  release includes a single-shot pilot (3 Claude models; 1,320 graded
  completions) and a pre-registered confirmatory multi-turn matrix (5 models
  across Anthropic, OpenAI, and Google; 3,250 graded completions).
keywords:
  - LLM agents
  - documentation drift
  - context rot
  - agentic benchmark
  - pre-registration
  - coding agents
license: CC-BY-4.0
version: "1.0.0"
date-released: "2026-06-16"
doi: 10.5281/zenodo.20722100
identifiers:
  - type: doi
    value: 10.5281/zenodo.20722100
    description: Concept DOI (always resolves to the latest version)