All research
Planned study · Results pending
Agent Invariant Preservation
A planned empirical study of how reliably coding agents preserve transactional invariants when modifying stateful backends.
Researcher
Research repositoryPythonAI AgentsSoftware CorrectnessTransactions
About this study
Preserving Transactional Invariants in AI-Assisted Software Development is a planned empirical study using the Dragon Market backend as its intended subject. It asks whether invariant-oriented guidance improves the correctness of changes made by coding agents. The repository contains the research protocol, draft task packages, provenance tooling, and analysis infrastructure. No empirical results or human-validated benchmark tasks are available yet; the intervention hypothesis remains untested.
Methods and research artifacts
- Controlled study design for coding agents on stateful backend tasks
- Draft tasks for purchase retries and bid cancellation retries
- Separate feature checks and transactional state assertions
- Research CLI with snapshots, source hashes, and execution provenance
- Explicit review and isolation gates before empirical agent runs