Skip to content

v0.7.0 — Conservative Dual-Verify for agent loops

Choose a tag to compare

View all tags
@azank1 azank1 released this 21 Jun 21:53
· 27 commits to main since this release
v0.7.0
23a0b51

Conservative Dual-Verify (CDV)

Agents submit step_output artifacts; the MCP server scores each step through two independent channels:

  • Channel A: deterministic evaluators (regex, JSON, completeness, composite)
  • Channel B: separate critic via MCP sampling (verifier hat)
  • Final score: min(channel_a, channel_b) — the stricter channel wins

Added

  • step_scorer.py, guards.py, evaluator_factory.py
  • Composable guard stack: timeout, token budget, output-repeat, plateau, threshold, Bayesian ROI
  • Async artifact-primary loopllm_loop_step; legacy score fallback with deprecation warning
  • Verdict JSON exposes channel_a_score, channel_b_score, score_source, deficiencies

Changed

  • AgentLoopController uses GuardStack
  • README restructured around three-layer architecture
  • Internal launch docs pruned; docs/launch/show-hn.md updated for CDV

Full changelog: CHANGELOG.md

Assets 2
Loading