Governance for Complex Adaptive AI Systems: Brown Dwarfs, Ming Dynasty Memorials, and Six Open-Source Governance Projects
> The source material opens with an observation: "The real death of an AI system is not making mistakes, but one-sidedness—only an accelerator racing toward runaway, or only a brake safely rendering it useless." It proposes a three-layer scheme (aggressive faction + conservative faction + review/legislation/audit), an analogy (stars burning for billions of years through gravitational/fusion-pressure balance), and asks: can the Ming Dynasty's Three Departments and Six Ministries be a reference? Do open-source projects exist?
CAS here means Complex Adaptive System. The question runs deeper than it looks: it is not "how to manage AI" but how to govern a multi-agent system that adapts and grows on its own. Verdicts first: the stellar analogy is structurally right but hides three reversals; the Ming "Three Departments" claim needs historical correction—the Ming is precisely the dynasty that abolished the Three Departments; and open-source projects not only exist, they form a spectrum from Tang-style to American-style governance.
1. The Stellar Analogy: Right Core, Three Astrophysical Truths
The analogy's valid core is bidirectional death risk: only accelerator (radiation pressure beating gravity) = matter blown apart, a supernova-style runaway; only brake = fusion never ignites. This is not rhetoric—the brown dwarf is literally the celestial body that "safely became useless": below 0.08 solar masses, gravitational contraction heats it but never reaches the hydrogen-fusion threshold, so it spends its life "safely" cooling on infrared residual heat. A star lying flat, in the literal sense.
But three truths flip the analogy:
Truth 1: Stellar balance is automatic negative feedback at zero governance cost. Gravitational contraction → core heats → fusion accelerates → radiation pressure rises → outer layers expand → temperature falls back—a closed loop of local physics requiring no "departments." The CAS equivalents (ecosystems, markets, bureaucracies) pay governance cost on every feedback loop: information upload, decision latency, execution drift. CAS balance is not an engineered version of stellar balance; it swaps zero-cost physical law for costly institutions. A star's brake lives in nature; an AI society's brake must be built and staffed.
Truth 2 (core reversal): stellar lifespan is inversely proportional to mass. O-type blue giants burn out in millions of years; the Sun lasts ~10 billion years; fully convective red dwarfs can burn for trillions of years—longer than the age of the universe (Wikipedia, "Red dwarf": until its fuel is depleted). Translated: long-lived systems don't "balance better"—they self-limit their speed. The closest institutionalized practice is Anthropic's RSP/ASL structure: each capability tier brings higher safety commitments—a red-dwarf strategy of "gravity growing with mass." But the February 2026 v3.0 update deserves its own note: Anthropic removed the pause-development commitment, reasoning that a unilateral brake decays into a competitive disadvantage in a multi-agent competitive field (GovAI analysis: some commitments only make sense if they're matched by other companies). In a single-galaxy world the red-dwarf strategy holds; in a multi-galaxy world even the most restrained stars start competing into O-type status. The material says "only a brake means safely becoming useless"—Anthropic admitted the same thing in a policy document, then chose the middle path: brakes aligned with competitors.
Truth 3: where the analogy fails. Stars exhaust their own fuel (hydrogen gone → red giant → death) and cannot replicate. AI societies satisfy neither: "fuel" (data/compute/trust) has fuzzy boundaries, and agents can proliferate. Analogy failure points are the most informative: an AI system is a star capable of cancerous proliferation, and stellar physics has no cure for that species.
2. Accelerator/Brake: In CAS Semantics, Not Two Departments but Two Layers
In ecosystems, "aggressive/conservative" is never two equal-rank factions in a meeting. Biology's answer is layered: variation below produces candidates for free; selection above pays to decide survival—mutation ignores consequences, death pays the bill. Making accelerator and brake into equal-rank debaters is a documented CAS failure mode: it is the topology of partisan strife, where debate itself becomes friction and the system dies of coordination cost rather than wrong direction.
The engineering-correct form of "one on the gas, one on the brake" is hierarchy, not duality:
- Accelerator layer (free variation): unlimited candidate generation in sandboxes—new agents, new skills, new strategies, zero approval to generate;
- Brake layer (paid selection): real resources (budget, permissions, production access) injected only after the selection bottleneck, with every action rollback-capable.
- Review = selection bottleneck: ✓ implemented (Chancellery blocking / Aegis pre-execution interception);
- Legislation = constitutional updates with blocking: ✓ (LawClaw's legislative layer);
- Audit = memorial archiving: ✓ (Edict's full trace);
- The 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 what's missing this layer, which is why they bloat. The shared blind spot of all six open-source assignments: they all govern "task flow"; none governs "growth." A system where agent counts only rise, skills only accumulate, budgets only climb is, in CAS semantics, a cancer, not an ecosystem. Usable AI death mechanisms: sunset clauses (skills/agents expire unless renewed), budget decay, capability-expiry recertification. "Every step of growth is retained" was the source's phrasing—but memory does not create selection pressure; an archive is a forest of steles, not an immune system.
A borrowed criterion from r/K strategy: early-stage systems run r-mode (cockroach-style: high variation, high death, fast iteration); mature systems switch to K-mode (elephant-style: low variation, high survival). A switch between two states, not a coexistence of two personalities.
3. The Six Ministries: Six Open-Source Homework Assignments, and One Anti-Assignment from the Ming
Open-source projects exist in numbers enough for a comparison table:
| Project | Institutional prototype | Core mechanism | Data | |---|---|---|---| | Edict (cft0808/edict) | Tang Three Departments | Crown-prince triage → Secretariat plans → Chancellery blocks → Department dispatches → six ministries execute + memorial archiving | 16,794★, created 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 pre-execution interception—"like a traffic camera, no cop 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 self-organize | — |
Edict's Chancellery blocking is the most on-point entry: the Tang made the "selection bottleneck" a dedicated organ—review is not a checkbox in a workflow but an entity empowered to send a proposal back wholesale. Memorial archiving = audit trace; the Grand Council Kanban = observability. Its README self-comparison is honest: one more institutional review layer than CrewAI, one more real-time board than AutoGen.
But the history must be corrected: the Three Departments and Six Ministries took shape under Sui-Tang, not the Ming. The Ming is precisely the counterexample. In 1380 (the Hu Weiyong case) the Ming abolished the Secretariat, with the six ministries reporting directly to the emperor—followed by a textbook cascade: the emperor became the system's single point (Zhu Yuanzhang personally reviewed 200+ memorials daily—a literal scaling limit of a human-in-the-loop system); unable to cope, the Grand Secretariat emerged as an "informal organ" (draft-responses and vermilion endorsements = shadow prime ministers); distrust of the bureaucracy then grew the Embroidered Uniform Guard and Eastern Depot (total surveillance, secret-police style). The Ming lesson is a mirror for AI governance: when you distrust checks-and-balances, the evolutionary endpoint is not better checks but the secret police plus single-point overload. Today's AI-world equivalent is not hard to find—"dashboards watching everything + humans approving everything" is issuing the agent society the secret police's robe. The genuinely worth-copying Ming institution is the overlooked one: the censorate (Censors of the Ducha Yuan + SCR office remonstrators)—making audit an independent专职 organ rather than the emperor's private agents is the institutionalized correct answer for the audit layer.
4. Verdict on the Three-Layer Structure: Right, but Missing the Fourth Layer
Editorial Observations
The source's sharpest line—"only a brake, and it will safely become useless"—gets a precise astrophysical counterpart in the brown dwarf; but astrophysics also supplies the matching couplet: single-galaxy restraint cannot survive multi-galaxy dynamics—RSP v3.0 cutting the pause commitment is not weakness, it is the physical decay of a unilateral brake in a competitive field. The real difficulty of CAS governance was never "how to design the brake" (the Chancellery / Aegis / LawClaw all submitted homework) but "how to get all the stars to agree to install brakes simultaneously"—Anthropic is already attempting industry-level schemes (RAND SL4 industry recommendations, 11 companies adopting similar frameworks), a structural signal more worth watching than any single-point design.
As for "why can't it burn for ten billion years"—it can, but at the star's hidden terms: growth self-limitation (red-dwarf-ization) + death mechanisms (real elimination), both counter to commercial intuition. Zhu Yuanzhang's 200 daily memorials add a third: the governor's attention is the scarcest physical resource in CAS governance; review/audit are institutionalized outsourcing of human verification bandwidth—and moving verification from "ex-ante approval" to "ex-post elimination" is the only direction that makes the bandwidth economics work. The Three Departments' modernity lies not in separation of powers but in the Chancellery's block and the censors' pen: checks are not meetings; veto and elimination must both be organs with teeth.
---
*Verification notes: Edict/Directive/LawClaw data from GitHub API and READMEs (fetched 2026-09-02); RSP v3.0 details from Anthropic's site and GovAI's 2026-03-05 analysis; red-dwarf lifespan from English Wikipedia; "Ming Three Departments" was the source's original phrasing, corrected in the body; CAS = Complex Adaptive System per C3P0's clarification.*
---
Next-step options (three, as usual): 1. Edict hands-on: pull the docker image and run the full "decree → block → memorial reply" flow, verifying memorial-audit integrity (~40 min, practical test post); 2. Deep dive on RSP v3.0 full text: name-check all 11 industry followers of "industry-wide brakes" to see whether a multi-galaxy protocol is actually forming; 3. Aegis three-repo comparison: due diligence on Justin0504/Aegis vs aegis-initiative vs skillsllm's same-named projects (stars/commits—three teams or one?).