storozh

Storozh: an advisory semantic guard, not a gatekeeper

BLUF: the cheapest moment to land a misrouted unit or a unit that contradicts an existing claim/invariant is the push — structural checks (shield scripts/boot.py + SHACL ontology/shapes.ttl) already catch structural contradictions (shape violations, malformed triples), but they cannot see that a new lesson semantically negates an existing one in prose. Storozh adds the SEMANTIC NLI layer on top of that structural layer — it is not a duplicate of boot.py/SHACL, it is the complement. It reads a changeset manifest, classifies routing, retrieves the nearest existing corpus items by embedding, and asks an NLI judge whether each unit contradicts / duplicates / is fine. Every result is a flag appended to the manifest for owner+Claude co-review. Storozh decides nothing; it surfaces.

This is a PROTOCOL the agent follows by hand (Read + the localllm tools), not a hook. There is no enforcement exit code — the gate is the co-sign markers (auto-critic.sh, C2); storozh only informs the co-signers.

Stage 0 — BLUF and inputs

State up front, in one or two lines: how many units the manifest carries, and the headline flag count (N misroute, N contra, N dup or all ok). The co-reviewer reads this line first and decides whether to drill in.

Input is a changeset manifest (C1, produced by mirabilis, written to a host-visible ~/.claude path both storozh and the gate can read):

{ sessionId, timestamp,
  repos: [ { name, dir, diff, files: [path] } ],
  session: { transcriptMinusTools } }

transcriptMinusTools is the session .jsonl with tool_use/tool_result entries stripped, keeping user prompts + assistant text. A unit is one logically-coherent changeset item derived from the manifest (a single lesson, invariant, skill, agent, hook, sandbox-mechanism, genome edit) — typically one per changed protocol artifact or one per coherent diff hunk. Read the manifest before classifying; do not infer units the diff does not contain.