CGP Reliability and Auditability Benchmark

Controlled benchmark suite for testing Continuity-Governed Prompting (CGP) against ordinary summary-based prompting in agent-assisted coding workflows.

Canonical repository target:

github.com/heart-ai-foundation/cgp-benchmark

Official manuscript DOI:

10.5281/zenodo.20234367

Publication date: May 16, 2026.

Public web anchors:

• Heart AI Foundation: https://heartaifoundation.org/
• HEART Standard v1.8: https://heartaifoundation.org/heart-standard/
• HEART Standard DOI: https://doi.org/10.5281/zenodo.20237387
• CGP paper DOI: https://doi.org/10.5281/zenodo.20234367

Epistemic Status

The registered primary endpoint was scope drift. The completed primary dataset showed baseline scope drift at the floor: 0 baseline scope-drift events in the preregistered primary dataset, 1 CGP scope-drift event, and an exact paired Wilcoxon two-sided p value of 1.0000. Across all planned primary plus companion-extension runs, scope drift occurred once under baseline and once under CGP. This benchmark therefore does not support a claim that CGP reduced scope drift in this run.

The operative observed findings are reliability and auditability findings: valid completion improved from 56 of 72 baseline runs to 68 of 72 CGP runs across all planned runs, work submission improved from 79.2% to 100.0%, and CGP evidence-trio completeness was 71 of 72. These findings must be read under the preregistration deviation note:

docs/research_integrity/CGP_Benchmark_Preregistration_Deviation_Note_v1_1.md

The registered analysis output is:

runs/processed/registered_analysis.md

Repository Layout

• benchmark-repo/ - controlled Python and JavaScript hybrid project used for task runs
• next-prompt-protocols/ - reference scaffold for CGP condition
• runs/raw/ - raw run logs, diffs, and telemetry exports
• runs/processed/ - cleaned run data
• scripts/ - reproducible analysis and harness scripts
• docs/paper/ - paper source, figures, tables, and references
• docs/ - experiment protocol and task documentation
• docs/osf_preregistration/ - OSF-ready pre-data preregistration packet
• docs/research_integrity/ - preregistration deviation and remediation notes
• docs/run_harness.md - post-preregistration execution procedure
• docs/prompt_templates/ - Baseline and CGP condition prompt templates
• runs/run_plan.csv - deterministic randomized run order generated after OSF preregistration
• scripts/prepare_run.py - prepares a single run prompt and optional isolated worktree

License

Code is licensed under MIT. Paper, protocol text, and data are intended for CC BY 4.0 publication unless a specific file states otherwise.