neuro-matrix

Advisory semantic guard over a changeset manifest before it lands in shared state. Use after a multi-repo session produces a changeset-manifest (mirabilis C1) and before the co-sign push gate — classify each unit's routing against the darwin routing-policy (C3) and NLI-check each unit against the shield corpus (claims + invariants + lessons + memory) for contradiction or duplication. Flags, never blocks — the owner and Claude co-review the flags and decide.

Run an independent audit gate before any fleet output lands. Use after agent-fleet-orchestration Stage 3 (or any multi-agent build) — before merge, before deploy, before the next phase. Frames the reviewer as an external adversary, not a helper; produces triaged findings with human triage for load-bearing blockers.

Run a lead + subagent fleet to build a non-trivial system without collisions, drift, or false completion. Use when a task is large enough to parallelize across agents — decomposing it, freezing contracts, driving execution, and closing each stage with real evidence, not agent reports.

The engineering invariant bar — what "good" means for systems built by this harness. Use as a checklist during design, as the audit criterion for adversarial-review, and as the source of linter/CI rules to mechanize. Not a style guide; these are correctness and safety properties.

Turn an observed harness deficiency or improvement idea into a landed, properly-gated change to the neuro-matrix protocol. Use when the user reports a harness gap or defect, proposes a protocol improvement, or asks to harvest improvements from a session or audit. Verifies the gap against current main first, then routes each change class to its correct gate.

Implement an academic paper (arXiv or any quantitative source) as verifiable code without inventing unstated details. Use when asked to implement, reproduce, or prototype a method from a paper. Produces an ambiguity audit before code, citation-anchored decisions, and machine-checkable sanity verification.