Trustworthy AI-Assisted Software Engineering
Ongoing controlled studies of memory in the AI coding technologies used in practice (coding agents, LLM-assisted code review, and verification pipelines) and in the models themselves: what degrades or distorts remembered experience, how it spreads between verifiers, how it should be handed over and when it must be re-verified, and how a fixed budget is best split between maintaining memory and verifying results. Builds on earlier work on measurement validity, residual failure after passing tests, unsafe artifacts, assurance gates, and temporal validity.
-
Defect Propagation Through Remembered Success
In everyday agent-assisted development, whether remembering a patch that passes its tests yet still hides a defect carries that defect into the next code an agent writes, examined through controlled reuse of past experience.
- Experience Reuse
- Defect Propagation
- Test-Passing Patches
-
Correlated Misses from Shared Review Records
In multi-reviewer code review pipelines, whether two AI verifiers that read the same past review record miss the same defect together, compared with verifiers that do not share a record.
- Multi-Agent Code Review
- Correlated Failure
- Review Memory
-
Handoff Format and Cross-Task Maintenance
When an agent carries past experience into a new maintenance task, whether handing the same evidence over as prose or as structured JSON changes task completion and regression defects.
- Handoff Format
- Cross-Task Maintenance
- Regression Defects
-
Code Layout and Defect Detection in Review
In LLM-assisted code review, whether arranging identical code and evidence by file path or by import relationship changes how accurately the reviewer points out real defects.
- Code Layout
- LLM Code Review
- Defect Detection
-
Specification Sufficiency in Code Review
In review of real change requests, whether an LLM can distinguish cases where the current candidate code can still be judged despite a missing requirement from cases where the specification must be clarified first.
- Specification Sufficiency
- Verdict Decidability
- LLM Code Review
-
Reverification After Code Change
In long-running development with memory-augmented coding agents, when a stored verification result should be invalidated and re-run once the code has changed, compared across reverification policies.
- Verification Staleness
- Reverification Policy
- Agent Memory
-
Memory Maintenance vs. Verify-and-Repair
Under the fixed compute and time budget of a real project, whether it is better to spend on organizing memory in advance or on verifying and repairing afterward.
- Memory Maintenance
- Verification–Repair
- Fixed Budget