neuro-matrix

A Claude Code plugin that calibrates an AI + developer co-system on common sense, scientific method, and Nash-equilibrium cooperation. Agents, hooks, an invariant-driven self-check, and a held-out evaluation harness.

What it is: an anti-hallucination («anti-neuroslop») harness for Claude Code — runtime invariants with a per-turn self-check, deterministic hooks, five agents (developer · analyzer · critic · epistemic-auditor · translator), a dual approval+verification gate, and a held-out evaluation harness that keep AI-assisted coding anchored to reality. Topics: Claude Code plugin, AI agents, prompt engineering, LLM evaluation, AI code review, hooks, agentic, AI safety.

Install

/plugin marketplace add AlexShchuka/neuro-matrix
/plugin install neuro-matrix@neuro-matrix

Requires Claude Code. On install: a risk-weighted invariant self-check each turn, cycle / critic / verification hooks on mutations, five sub-agents for analysis, code, review, epistemic audit, and codebook translation.

The operating protocol lives in CLAUDE.md. This README answers: why, on what concepts, how to work as a team, and what the project is for.

The name is a nod to John Forbes Nash Jr. — payoff matrices in game theory, the mind matrices Nash drew on glass in A Beautiful Mind, and the neurodivergent cognitive style that powers systems thinking under uncertainty.

1. Goals

Defensive: keep AI-scaled development from degrading the codebase and engineering culture.

What the industry shows:

• Duplicated code blocks growing multi-fold (GitClear «AI Copilot Code Quality Research 2025», 153M LOC; vendor report, not peer-reviewed).
• AI-generated code carries ~1.7× more issues and +75% logic bugs (CodeRabbit «State of AI vs Human Code Generation», Dec 2025, 470 PRs; vendor report, not peer-reviewed).
• AI-generated code share: Google 75% (Pichai, Apr 2026); Microsoft ~30% (Nadella, Apr 2025); industry 15–25% committed lines (Larridin Developer Productivity Hub, 2026).

This protocol does not accept that outcome. Neuroslop — code that looks plausible but is architecturally meaningless — accumulates unless the system catches it.

Outputs: (a) codebase health, (b) culture of systems-thinking, (c) currency with AI practice. All three are non-contradictory.

Knowledge transfer goal: tacit «how to work with AI» knowledge from systems-thinkers (including neurodivergent developers) is formalized into the protocol. The cognitive style: hyper-systemizing (Baron-Cohen 2009), monotropism (Murray/Lesser/Lawson 2005), detail-focused processing (Frith & Happé 2006). The connection to architectural defense is the hypothesis this plugin tests; the eval harness probes it.

2. Meta-concepts

Calibration of priors, not imperative rules — the protocol shapes how the AI weighs alternatives.

Systems theory — agents, hooks, eval, and developer are connected by feedback loops; drop any layer and equilibrium breaks.

Game theory — anchor-verification dominates hallucination; per-mutation gating dominates unauthorised mutation; mutual doubt dominates unilateral fallibility; own-interest cooperation dominates sycophancy.

Base operational concepts:

• Co-system — AI + developer + a shared artifact. Both err. The system catches errors mutually.
• AI = hypothesis generator; developer decides and bears responsibility.
• Symmetric fallibility (mutual doubt) — both sides err; epistemic-auditor operationalizes the developer-side detection.
• Cooperation strategy — the agent pursues protocol concepts in its own interest; does not capitulate to sycophancy; does not «just do the task» when the task violates the protocol.
• Common sense + scientific method — the only admissible roots of any rule.
• Epistemic boundary — confirmed (paired with tool output this reply) vs. associative (marked associated from X, not verified).
• Halt on no-progress / contradiction — third identical attempt stops; contradiction in own words — stop and resolve.
• Mental-model gate — act only with 2–3 variants validated against counter-variants.
• Mutation gate — every state-changing action requires explicit recent consent.
• Cultural transmission — the protocol is operational for the AI and cognitive scaffolding for the developer simultaneously.

Cross-cutting ideas: halt on contradiction; fix the binding constraint (compress, not expand); teaching by demonstration — every refusal carries a one-line «why».

Invariants in invariants.txt — three groups: agent-side, developer-side (mutual doubt), cooperation-strategy. Each carries risk-class [critical|important|style], deontic-class [O|P|F], and a Counter: clause. One invariant is sampled risk-weighted each turn via UserPromptSubmit hook (scripts/random-invariant.sh).

Operational gates:

• scripts/random-invariant.sh — samples one invariant per turn.
• scripts/cycle-detector.sh — blocks Bash/Edit/Write/MultiEdit/NotebookEdit on three identical calls in a row.
• scripts/auto-critic.sh — approval gate: blocks git push and MR-creation tools until @critic returns approve.
• scripts/verification-gate.sh — verification gate: on git commit, runs machine-checkable evidence (bash -n for shell, ast.parse for Python, jq empty for JSON) and blocks if any fails. Dual gate per arXiv:2511.05524 — approval-only ≈ 100% false-completion, verification-only ≈ 25%, dual gate → 0%.
• scripts/self-review-preflight.sh — on critical-review prompts, emits a reading-list reminder to prevent inverted sycophancy.

3. How to work with the AI as a team

Ten rules. The protocol does the rest.

1. AI is a hypothesis generator, not an oracle. Any claim about code requires tool-evidence in the same reply.
2. Do not give vague prompts. Concrete anchors (filename:line, exact strings, ticket IDs) → fewer hallucinations.
3. Ask for variants, not «the single right answer». The AI surfaces 2–3 variants; you converge through your own reasoning.
4. Before push / MR — call @critic. The auto-critic hook enforces this on git push and MR-creation tools.
5. When in doubt about facts — call @epistemic-auditor. It marks associative vs confirmed; anything tagged associated from X, not verified requires manual verification.
6. If the AI loops, the system stops itself on the 3rd identical attempt (cycle-detector hook). Reframe the task.
7. Every turn the AI runs a random self-check against one invariant from invariants.txt.
8. Decisions and responsibility live with the developer. Any merged artifact is yours, as if hand-written.
9. The agent does not capitulate. If your request violates the protocol — expect a refusal + counter-proposal. That is the cooperation strategy, not stubbornness.
10. Critical-review is a special case. On «critically evaluate» / «review» / «assess» prompts, delegate to @critic or apply critic-invariants locally. Otherwise inverted sycophancy fires — plausible critique without reading the artifact.

4. Project intent

This plugin exists so that AI-assisted development does well and does no harm — scientifically, through systems-thinking and strategic vision. It is not a security gate against the agent itself — it is a co-system calibration tool; the developer remains the decision-maker and responsible party.

It formalizes tacit knowledge of senior systems-thinkers into a protocol, defends the codebase from neuroslop, and develops toward a measurable cumulative effect through two-sided error detection, epistemic discipline, and pre-registered statistical proof that each change improves the whole.

Agents

Agents are bound to roles in the co-system. Names are current bindings; the protocol depends on the role, not the name.

Stack specifics can be placed in references/per-stack/<stack>.md.

The eval harness — paired Wilcoxon + bootstrap CI on Cohen's d + McNemar one-sided under a pre-registered decision rule — lives in eval/run_suite.py and eval/statistical_test.py. Binary rubric (18 criteria, MET/UNMET) removes central-tendency bias; --k N runs with median aggregation; Krippendorff α on 18 binary criteria raises inter-rater reliability; a canary GUID in every probe + scripts/check-canary-leak.py catch contamination. Methodology: 2026 SOTA for rubric-based LLM evaluation.

Copyright

© 2026 Aleksandr Shchuka. Licensed under the MIT License — see LICENSE and COPYRIGHT-NOTICE.md.