Hermes Blind 0.2.0 adds a machine-readable Reliability Lab result envelope and structured recovery-anchor observability without changing existing extraction semantics.
Highlights:
python -m hermes_blind.evidenceemits a JSON result for an explicit session path.build_anchor,AnchorResult, andParseStatsexpose extraction facts without parsing rendered markdown.- Unparseable lines, unmatched sentences, and ambiguous initial turns are reported rather than silently inferred.
- Session paths remain explicit and are reported by basename only.
The envelope reports deterministic parsing and extraction results; it does not detect drift or establish downstream model recovery.