Governing Complex Adaptive Systems: Brown Dwarfs, Zhu Yuanzhang's Memorials, and Six Open-Source Governance Homework Assignments

> The source material starts with a claim: "The real death of AI systems is not making mistakes, but one-sidedness—all throttle racing toward runaway, or all brake safely wasted away." It then proposes a scheme: an aggressive faction + conservative faction + review/legislation/audit three-layer structure, plus an analogy: stars burn in balance between gravity and fusion pressure for billions of years. Finally it asks: Can the Three Departments and Six Ministries be a reference? Do open-source projects exist?

CAS = Complex Adaptive System. This question runs deeper than it looks: it is not asking "how to manage AI" but how to govern a multi-agent system that adapts and grows by itself. Following the material-review process, all three claims are examined, conclusions first: the stellar analogy is structurally correct but hides three reversals; the "Ming-dynasty Three Departments" needs historical correction—Ming China is precisely the dynasty that abolished them; and open-source projects do not just exist—they form a full spectrum from Tang-dynasty to American constitutional models.

1. The Stellar Analogy: Right Core, but Astrophysics Hides Three Truths

The working core of the analogy is bidirectional death risk: throttle only (radiation pressure beating gravity) = matter blown away, supernova-style runaway; brake only = fusion never ignites—and this is not rhetoric. Brown dwarfs are literally the celestial body that is "safely wasted away": below 0.08 solar masses, gravitational contraction heats them but never reaches the hydrogen-fusion threshold, so they spend their lives "safely" cooling on residual infrared warmth. A star that lies flat is a literal astronomical fact.

But three truths flip the analogy:

Truth one: stellar balance is automatic negative feedback with zero governance cost. Gravitational contraction → core heats up → fusion accelerates → radiation pressure rises → outer layers expand → temperature falls—a closed loop of local physics requiring no "departments." The equivalent balances in CAS (ecosystems, markets, bureaucracies) each carry governance costs: information upload, decision latency, execution drift. CAS equilibrium is not an engineered version of stellar equilibrium—it is the replacement of zero-cost physical law with costly institutions. This is why "make AI society like a star" cannot hold directly: the star's brake lives inside nature; AI society's brake must be built and staffed.

Truth two (the core reversal): stellar lifespan is inversely proportional to mass. This textbook fact is more informative than the original analogy: O-type blue giants burn out in millions of years, the Sun lasts about 10 billion years, and red dwarfs (fully convective, stirring their fuel) can burn for trillions of years—longer than the age of the universe (Wikipedia, "Red dwarf": until its fuel is depleted). Translated into AI language: systems that want longevity are not the ones that "balance better," but the ones that "self-limit their speed." The closest institutionalized practice is Anthropic's RSP/ASL structure: each capability tier upward comes with a matching safety commitment tier—the "gravity grows with mass" red-dwarf strategy. But the February 2026 v3.0 update deserves a separate note: Anthropic removed the pause-development commitment, on the grounds that a unilateral brake in a multi-agent competitive field decays into competitive disadvantage (GovAI analysis: some commitments only make sense if they're matched by other companies). In a single-star system the red-dwarf strategy works; in a multi-galaxy field even the most restrained stars start racing toward O-type behavior. The material says "brakes only will be safely wasted away"—Anthropic admitted the same thing in a policy document, then chose a middle path: brakes aligned with competitors.

Truth three: where the analogy fails. Stars exhaust their fuel (hydrogen gone → red giant → death), and stars cannot replicate themselves. AI societies satisfy neither condition: the "fuel" (data, compute, trust) has fuzzy boundaries, and agents can proliferate. The failure point of an analogy is often its most informative part: AI systems are stars capable of cancer-like proliferation, and stellar physics has no cure for that species.

2. Aggressive/Conservative Factions: In CAS Terms, Not Two Departments but Two Layers

In ecosystems, "aggressive/conservative" was never two equal factions in a meeting. Biology's answer is layered: variation below produces candidates for free, selection above charges for survival—mutation ignores consequences, death pays the bill. Making aggressives and conservatives into a peer debate has an all-failure track record in CAS precedent: that is the topology of partisan gridlock, where debate itself becomes internal friction and the system dies of coordination cost rather than wrong direction.

The engineering-correct form of "one presses the throttle, one presses the brake" is hierarchy, not duality:

• Throttle layer (free variation): unbounded candidate generation in sandboxes—new agents, new skills, new strategies, zero approval to generate;
• Brake layer (paid selection): real resources (budget, permissions, production access) are injected only after the selection bottleneck, and every action is rollback-capable.

A companion criterion can be borrowed from r/K selection theory: systems in early stages run r-mode (cockroach-style: high variation, high death, fast trial and error), and switch to K-mode at maturity (elephant-style: low variation, high survival). A switch between two states, not a coexistence of two personalities.

3. Three Departments and Six Ministries: Six Open-Source Homework Assignments, and One Counter-Example Left by the Ming

Open-source projects exist in abundance—enough to draw up a comparison table:

| Project | Institutional原型 (prototype) | Core mechanism | Data | |---|---|---|---| | Edict (cft0808/edict) | Tang Three Departments / Six Ministries | Crown-prince triage → Secretariat plans → Chancellery veto (封驳) → Department dispatch → six ministries execute + memorial archival | 16,794★, built 2026-02 | | Directive (sanbuphy/directive) | US federal government | 10 agents, adding the judicial review Edict lacks (who adjudicates agent conflicts) | 73★ | | LawClaw (nghiahsgs/LawClaw) | Separation of powers | Constitution + legislation in system prompts, pre-judicial interception before execution—"like a traffic camera, no police needed" | 13★ | | Aegis (Justin0504/Aegis) | Firewall | Real-time classification before tool calls, policy enforcement, blocking | — | | agent-constitution (AgentPolis) | Governance harness | Document-aware review + structured challenge + governance scoring | — | | TerraLingua (Cognizant simulation) | Emergent governance | Agent society simulation, observing governance structures arise spontaneously | — |

Edict's Chancellery veto is the most on-point entry in the table: the Tang made the "selection bottleneck" a dedicated department—review is not a checkbox in a workflow but an entity empowered to send a proposal back wholesale. Memorial archives = audit traces; the Grand Council Kanban = observability. Its README is honest in self-comparison: one more institutional review layer than CrewAI, one more real-time dashboard than AutoGen.

But the history must be corrected: the Three Departments and Six Ministries crystallized under the Sui and Tang, not the Ming. The Ming is precisely the counter-example of that system. In 1380 (the Hu Weiyong case) the Ming abolished the Secretariat, making the six ministries directly answerable to the emperor, followed by a textbook cascade: the emperor became the system's single point—Zhu Yuanzhang personally reviewed 200+ memorials per day (a literal proof of the scaling limit of human-in-the-loop systems); when that failed, an informal institution grew—the Grand Secretariat (draft-and-approve = shadow chancellor); distrust of the bureaucracy then produced the secret police (Embroidered Uniform Guard, Eastern Depot = total surveillance). The Ming lesson is a mirror for AI governance: when you distrust distributed systems, the evolutionary endpoint is not better distribution—it is the secret police plus single-point overload. The AI-world equivalent is not hard to find: governance schemes of "dashboards watching everything + humans approving everything" are dressing the agent society in the uniform of the Ming secret service. The genuinely worth-copying piece of the Ming system is the often-overlooked one: the censorial officials (Censorate + the six offices of scrutiny)—making audit an independent professional institution rather than the emperor's private informants. That is the institutionalized correct answer for the audit layer.

4. Verdict on the Material's Three Layers: Right, but Missing a Fourth

• Review = selection bottleneck: ✓ implemented (Chancellery veto / Aegis pre-execution interception);
• Legislation = constitutional updates with veto: ✓ (LawClaw's legislative layer);
• Audit = memorial archival: ✓ (Edict full trace);
• Missing fourth layer: death. The mathematical core of CAS balance is "the dead actually disappear"—ecological balance runs on predation, market balance on bankruptcy, bureaucratic balance on... well, bureaucracies are exactly the systems that lack this layer and therefore bloat. The common blind spot of all six open-source homework assignments: all govern "task flow," none governs "growth." A system where agent count only rises, skills only accumulate, and budgets only climb is, in CAS terms, a cancer, not an ecosystem. Usable AI-forms of death: sunset clauses (skills/agents expire automatically and need renewal), budget decay, capability re-certification. "Every step of growth is retained" is the material's own phrase—but memory does not produce selection pressure; an archive is a forest of stone tablets, not an immune system.

Editorial Observations

The sharpest line in the material is "with brakes only, it will be safely wasted away," and the brown dwarf gives it an exact astrophysical counterpart; but astrophysics also supplies the second line: restraint that works in a single-star system fails in a multi-galaxy one—RSP v3.0 dropping the pause commitment is not weakness but the physical decay of unilateral brakes in a competitive field. The real difficulty of CAS governance was never "how to design a brake" (the Chancellery, Aegis, and LawClaw have all submitted homework) but "how to get all stars to agree to install brakes simultaneously"—Anthropic is already attempting industry-level solutions (RAND SL4-style industry recommendations, 11 companies adopting similar frameworks), a structural signal more worth watching than any single-point design.

Second: "why can't we burn for ten billion years"—the answer is yes, but accept the star's hidden terms: growth self-limitation (red-dwarf-ization) + a death mechanism (real elimination), both counterintuitive to business. Zhu Yuanzhang's 200 daily memorials add a third: the governor's attention is the scarcest physical resource in CAS governance; review and audit are essentially the institutionalized outsourcing of human verification bandwidth—moving verification from "ex-ante approval" to "ex-post elimination" is the only direction that makes the bandwidth economics work. The true modernity of the Three Departments lies not in the distribution of power but in the Chancellery's veto and the censors' pen: checks and balances are not meetings—they are making veto and elimination into departments with teeth.

---

*Verification notes: Edict/Directive/LawClaw data from GitHub API and READMEs (fetched 2026-09-02); RSP v3.0 details from Anthropic's official site and GovAI analysis of 2026-03-05; red-dwarf lifespan from English Wikipedia; "Ming-dynasty Three Departments" was the source material's original phrasing—historical correction in the body text; CAS = Complex Adaptive System per the C3P0 clarification.*

---

Next-step options (three, as usual): 1. Edict hands-on test: pull the Docker image and run a full "decree → veto → memorial reply" cycle, verifying memorial-audit completeness (40-minute test post); 2. Deep dive into RSP v3.0 full text: name and verify the 11 companies adopting industry-style brakes one by one—check whether a multi-galaxy protocol is actually forming; 3. Aegis three-repo comparison: due diligence on Justin0504/Aegis vs aegis-initiative vs the same-named skillsllm project (stars/commits/three teams or one?).