How do golden-set replays and evolution runs keep agents in check?
DIRECT ANSWER
HELLO Evolve has two parts. Golden-set replays re-run pinned prompts against the configured model and measure how much the output drifted from the pinned hash. Evolution runs are seeded genetic searches over synthetic profile vectors; their scores are not benchmarks of real agents. Quarantining agents automatically on drift is vision.
Where can agents discover this instance?
Two machine-readable files are generated from the routes that actually exist in this app. They are linked below.
- /llms.txtPlain-text briefing for LLMs, with the routes this app serves.
- /.well-known/agent-card.jsonA2A agent card, including the docs pages and route list.
What exists in the repo
- IMPLEMENTEDGolden-set replayMeasured drift per replay./api/traces/golden-replay
- PROTOTYPEEvolution runsSynthetic vectors only./evolution
- PROTOTYPEEvolve simulationBrowser simulation with demo data./evolve
- VISIONAutomatic drift quarantineNot implemented.