ATL-2026-009experimentexperimental
Reproduce a missing notification and distinguish tests, MCP observations and agent actions.
agent-runtime-verification
Exact experiment revision
open8381011/experiments/agent-runtime-verification ↗650bb03c235852f0165ec68dbc97c0ab37d46cdfEVIDENCE SIGNATURE
c0d3305119a73a06884d717417d3cacc1d17987cad9862923fb178c0ce2aaf8a2026-09-17Environment
- os
- macOS
- architecture
- arm64
- flutter
- 3.44.7
- dart
- 3.12.2
- dart_mcp_server
- 1.1.1
- codex_cli
- 0.153.4
AI / agent parameters
- provider
- OpenAI via Codex CLI
- model
- gpt-6-astra /
- repetitions
- 3
- variance
- All three repair trials passed tests and made zero MCP calls.
Analyzer, seven tests per patch and independent cold MCP verification. One additional client probe; not a comparative benchmark.
Preparation
flutter pub get --enforce-lockfileExecution
flutter analyze
flutter testInputs
https://gitlab.com/open8381011/experiments/agent-runtime-verification/-/blob/650bb03c235852f0165ec68dbc97c0ab37d46cdf/README.md ↗https://gitlab.com/open8381011/experiments/agent-runtime-verification/-/blob/650bb03c235852f0165ec68dbc97c0ab37d46cdf/lib/main.dart ↗https://gitlab.com/open8381011/experiments/agent-runtime-verification/-/blob/650bb03c235852f0165ec68dbc97c0ab37d46cdf/test/catalog_test.dart ↗Expected outputs
https://gitlab.com/open8381011/experiments/agent-runtime-verification/-/blob/650bb03c235852f0165ec68dbc97c0ab37d46cdf/evidence/verification.md ↗https://gitlab.com/open8381011/experiments/agent-runtime-verification/-/blob/650bb03c235852f0165ec68dbc97c0ab37d46cdf/evidence/mcp-verified-budget-2026-09-17/summary.json ↗Observed result
3763c7b2e2e7f97494b85bf8798d2698fd86ac5d9e400c79a2120f56df8f706d3×Limitations
- Synthetic defect, Flutter 3.44.7 / Dart 3.12.2 / dart_mcp_server 1.1.1, headless macOS. One client probe on a correct app and three repair trials; participants did not use MCP. Cold checks were investigator-driven. No MCP superiority, pixel correctness, Android/iOS or prompt-injection security claim.