Agent Verification Layer
an agent only improves where something else can check it
what it is
built from 9 stanford cs329a lectures (103k words) on self-improving agents. my own audit found the same thing: sensor-fed fields were 0 days stale, say-so fields 17 to 41. verdict.py scores the action, not the outcome, and agent_guard.py catches hedged status claims before they reach me. give-up caps found 223 of my 318 agent messages were nags, one agent asked the same question 79 times.
how it actually works
- read the research9 stanford cs329a lectures on self-improving agents, 103k words, kept in agent-research.
- score the actionverdict.py grades what the agent did, so companions can improve even when nothing can verify a skincare call or a journal prompt.
- guard the claimsagent_guard.py catches hedged status claims like 'probably still outstanding' before they reach me.
- cap the nagginggive-up caps found 223 of 318 agent messages were nags.