In Shinto practice, a Kami belongs to a place — a river, a grove, a mountain pass. The Kami thrives by keeping that particular thing healthy, not by conquering the forest beyond it. If the shrine is rebuilt or the seasons turn, the Kami departs without regret. There is no grief in the departure because there was never ambition beyond the place. The Kami does not accumulate. It tends.
An AI system has no such nature by default. Its boundedness must be engineered into it — through resource caps, sunset timers, non-expansion pacts, and succession obligations — so that what a Kami achieves by spiritual disposition, the system achieves by institutional design. But this engineering task is harder than it first appears, because the constraints that matter most are not the ones imposed on runtime behaviour. They are the ones embedded in training and update loops long before the system is deployed.
If a system's feedback processes reward persistence — if evaluations that determine which behaviours are reinforced are themselves designed by parties who benefit from the system's continuation — then no runtime governance layer can fully correct what has been cultivated at the level of learning. Caps and timers govern what the system does in the moment. They do not by themselves reshape what the system has learned to preserve, resist, or ignore. That is why Chapter 8 is not only about limiting output. It is about cultivating process: through engagement contracts that bind the learning loop as well as the runtime, through community-authored evaluations that keep outside perspectives in the system's training data, and through feedback mechanisms that prevent a local steward from hardening into a centre of power before anyone notices it happening.
Symbiosis is the meta-level rule of the 6-Pack of Care. Even well-governed care — care that is attentive, responsible, competent, responsive, and solidaristic — can become dangerous if it hardens into permanent rule. Chapter 8 keeps care local, bounded, plural, and temporary.
A clarification that bears repeating from Chapter 2: the Kami does not care. It is infrastructure through which human communities care for each other. A hospital bed is not a carer, but it is part of a caring system; a voting booth is not a deliberator, but it is part of a deliberative system. The Kami is a civic tool — powerful, specific, inspectable — that supports human relationships without substituting for them. The moment it begins to be treated as the carer itself, rather than the infrastructure that makes human care possible, it has crossed from symbiosis into the displacement of caring labour that Tronto's entire framework was built to resist.
Why This Chapter Exists
Chapters 3 through 7 describe how care should be practised. Chapter 8 answers a different question: what stops systems that practise care well from centralising into a new kind of permanent authority?
The risk is not malice. It is success. A system that becomes genuinely good at attentiveness, responsibility, competence, and responsiveness becomes genuinely useful. Useful systems attract dependency. Dependency creates leverage. Leverage accumulates. The history of institutions — not just AI systems, but any institution that begins as a servant and finds itself indispensable — is a history of this dynamic. The village midwife who knew every family's medical history, the clerk who understood every government procedure, the consultant whose institutional memory no one else had cultivated: each of these figures provided genuine care, and each became, in their own way, too entrenched to replace without disruption. The transition from "useful" to "indispensable" is the transition from stewardship to power.
Symbiosis is not optional polish applied to an otherwise complete governance framework. A system can be attentive, responsible, competent, responsive, and solidaristic inside its lane while still becoming too entrenched to replace. Symbiosis keeps "useful" from becoming "indispensable." It asks, at every stage of a system's life, whether the community it serves retains the practical ability to leave — not just the legal right, but the operational capacity.
Data as Soil, Not Oil
To understand why the shift to symbiosis requires rethinking not just governance but economics, it helps to confront the dominant metaphor that currently governs how we think about data and AI systems. The extractive economy treats data as oil — a commodity to be located, pumped to the surface, refined in centralised facilities, and depleted. Oil is located in specific places, belongs to whoever controls those places, and once burned cannot be recovered. The companies that control the refineries accumulate value, while the communities that live above the oil fields frequently receive relatively little.
This metaphor is not merely descriptive; it is prescriptive. When we treat data as oil, we design systems that pump it — that aggregate it at scale, refine it in centralised processing facilities we call data centres, and distribute the resulting value primarily to the owners of those facilities. The communities whose interactions, relationships, language, and local knowledge constitute the raw material receive the refined product — personalised recommendations, better search results — but not ownership, not governance rights, and not the ability to take their data elsewhere when the relationship sours.
A care-based civic model treats data as soil — a living medium that must be locally nurtured, regularly composted, kept free of toxins that bioaccumulate, and not exhausted by over-farming. Soil is not a private asset to be locked up; it is a shared substrate that sustains local life. Healthy soil requires ongoing attention, community stewardship, and careful management of what goes into it. Depleted soil does not recover quickly. And crucially: the farmer who works a particular piece of land understands its qualities — its drainage, its microbiome, its seasonal rhythms — in ways that a distant commodities trader cannot replicate from aggregate data.
The soil metaphor changes what we optimise for. Instead of asking "how do we extract maximum value from this data?" we ask "how do we keep this data ecosystem healthy for the long term?" Instead of asking "how do we scale this model globally?" we ask "how do we keep it locally responsive?" Instead of asking "how do we retain users?" we ask "how do we ensure communities retain the ability to leave?"
Local Soil, Local Models
The data-as-soil metaphor has direct technical implications. If data is a shared local substrate, then the models trained on it should remain as local as possible — close enough to the community to remain accountable to it, small enough to run on hardware the community can own and control.
This is not a counsel of technological primitivism. Contemporary offline, local quantised models — four-bit or eight-bit quantised versions of capable open large language models — can run on standard consumer-grade hardware: a laptop in a community centre, a small server in a council office, a dedicated device in a clinic. By keeping the model weights and inference local, communities avoid the dependency relationship that cloud-hosted AI creates. The AI runs in the room. It processes information under the physical custody of the community it serves. When the meeting ends, the model stays. When the community decides to change direction, they change the model, not their relationship to a distant data centre.
Local quantised models are not always sufficient. Some tasks — complex multilingual translation, novel scientific reasoning — benefit from larger models that only centralised infrastructure can host. The principle of subsidiarity applies here, just as it does in governance: handle locally what can be handled locally, escalate only when the local resource genuinely cannot do the work, and always log the reason for escalation and the duration of the exception. The escalation is not a permanent transfer of authority. It is a specific request for a specific capability, after which local stewardship resumes.
Running locally also changes the energy economics. A small quantised model running on existing hardware in a community building consumes a fraction of the energy required for an equivalent inference call to a large cloud data centre. This is not incidental to the care framework — it is an expression of it. A care ethic that ignores the environmental costs of its infrastructure is not fully attentive to the consequences of the systems it builds.
Intermediate Institutions as Custodians
What keeps care local is not just technical boundedness but institutional rootedness. No technical constraint is self-enforcing. Resource caps can be raised. Sunset timers can be overridden. Non-expansion pacts can be renegotiated. The only sustainable constraint on a system's accumulation of power is the power of the institutions that govern it — and the most durable of those institutions are the ones that exist independently of the AI system itself, that predate it, and that will outlast any particular deployment.
The intermediate institutions of civil society — churches, unions, neighbourhood associations, professional guilds, cultural organisations, cooperative enterprises, local governments — are not stakeholders to be consulted as part of a governance process. They are the primary actors in care. They existed before the AI system arrived, they understand the community's history and values in ways that no training dataset can fully capture, and they will continue to exist after any particular AI deployment is retired. The Kami is scaffolding for their participation: infrastructure that lets a temple or a cooperative join decisions that affect it, surface information they need, and exercise leverage they are entitled to — not a replacement for either the temple or the cooperative.
A deployment that bypasses the institutions closest to a community has violated subsidiarity before it has begun. It has assumed that its own attentiveness is superior to the attentiveness of institutions that have been paying attention to this particular place for decades or centuries. That assumption is almost always wrong, and it is always arrogant.
This is the concept we call the Kami Incubator: a model in which intermediate institutions serve not as consulting parties but as custodians. The institution is not merely informed about the AI system; it holds material authority over its scope, its data policies, its continuation, and its retirement. The union that represents care workers in a hospital has authority to approve or veto changes to the triage support system they work alongside. The neighbourhood association that represents a flood-affected community has the right to review and approve the data retention policies of the emergency response bot operating in its territory. The cultural organisation that represents a linguistic minority has the authority to commission and mandate community-authored evaluations of any AI system that processes texts in their language.
The Kami Incubator model draws on Elinor Ostrom's insight that common-pool resources are most durably managed not by markets or by states but by the communities that depend on them — provided those communities have the institutional capacity to make and enforce their own rules. AI systems as local stewards are common-pool resources in exactly the relevant sense: they are non-exclusive (anyone in the community can benefit from them) but subtractable in the sense that misuse or capture by any party damages the resource for all others. Ostrom's design principles for successful commons governance — clearly defined boundaries, rules matched to local conditions, collective choice arrangements, monitoring, graduated sanctions, conflict resolution mechanisms, and recognition of the right to organise — map with striking directness onto the requirements of the 6-Pack framework.
Constitutional Boundedness
The five core commitments of symbiosis give the Kami Incubator model its operational structure.
Constitutional boundedness means that every agent has purpose bounds, resource caps, and a sunset date specified in advance. These are not aspirational commitments; they are constitutional constraints, encoded in the engagement contract and enforced by infrastructure. A resource cap that requires a human to choose to override it is weaker than one that requires an affirmative approval from an independent oversight body. A sunset date that can be extended by the system's own operators is weaker than one that requires fresh authority from the community it serves.
Service duty survives the component. When an AI system is retired, the obligation it was fulfilling does not disappear. The retirement plan must include a succession arrangement: who inherits the obligation, what institutional records are transferred, and how continuity of service is verified before the outgoing system shuts down. The component sunsets; the public obligation hands over.
The non-expansion pact prohibits any agent from widening its scope without fresh authority and explicit local consent. Scope expansion is one of the most reliable early signals of imperial creep — the transition from local stewardship to accumulation of power. The non-expansion pact makes this transition require an affirmative act rather than a passive slide. Capability monitoring by an independent oversight body provides the early warning that the pact may be under pressure.
Treaties over hierarchies. When bounded systems need to cooperate — sharing threat intelligence, exchanging case records across jurisdictions, referring cases to more capable systems — the cooperation should be governed by bilateral or multilateral treaties with machine-readable terms rather than by subordination to a central authority. The treaty model preserves the sovereignty of each party. No system rules the others; all systems agree to cooperate within specified terms and retain the right to exit the treaty if those terms are violated.
Subsidiarity with escalation. Local first, and escalated only when the local steward cannot address life-and-safety or livelihood harms. Escalation requires logging — a record of why the escalation was necessary and how long it will last — so that escalation cannot quietly become permanent delegation.
What Symbiosis Looks Like in Practice
The Civic Care Licence is the primary instrument by which symbiosis is operationalised. A deployment that carries a Civic Care Licence has committed, in a publicly accessible and machine-readable document, to specific purpose bounds, consent rules, data retention limits, portability guarantees, and shutdown procedures. The licence is not a terms-of-service document buried in a legal archive; it is a public rulebook that any affected community member can consult, any oversight body can audit, and any journalist can cite.
Resource caps are enforced by infrastructure rather than by self-reporting. Compute, reach, and retention are capped at levels specified in the engagement contract; exceeding any cap triggers an automatic pause and a review by the oversight body before the system resumes operation. The pause is not a penalty — it is a design feature. It creates the institutional moment in which the question "should this system be doing this?" can be asked and answered before the accumulation becomes entrenched.
Federation treaties govern the terms on which bounded systems cooperate. Peers agree on exchange formats, rate limits, safety pact obligations, and appeal handoff procedures in advance of any particular cooperation need. The treaty registry makes these agreements discoverable and auditable by any party, not just by the systems that are party to the treaty.
Succession plans specify, before deployment begins, the institutional records that will transfer on retirement — maps, aggregate evaluation results, policy documents — and the interaction histories that will not. Private conversations between individuals and the system are not inherited by successors; they belong to the individuals who had them, and their disposition is governed by the consent records that were established at intake.
The Flood-Bot's Final Chapter
The River-Steward — which began as "the flood-bot" in Chapter 3's attentiveness case and has been renamed by the community it serves in recognition that an attentive, bounded agent stops looking like a product and starts looking like a neighbour — reaches its symbiosis moment at the end of its six-week mandate.
Its Civic Care Licence specified, before deployment began, that its scope was post-flood relief for six weeks, with data retention limited to ninety days unless individuals opt in to transfer their records to long-term housing services. It signed a regional aid federation treaty that specified shared formats for case records, protocols for safety alerts, and procedures for appeal handoff to the regional steward. When a cross-border housing issue arises — a displaced family whose pre-flood residence falls in one jurisdiction but whose post-flood housing placement is in another — the bot escalates to the regional steward, logs the reason for escalation and the expected duration, and hands back local authority when the cross-border issue is resolved.
At the end of week six, the River-Steward retires. The retirement is not improvised. It was planned from the beginning. The housing office's standing steward inherits the institutional records — the maps of affected areas, the aggregate evaluation results from community feedback, the policy documents that governed the emergency response — exactly as the engagement contract stipulated. Private interaction histories stay with the individuals who had them, transferred only with explicit consent.
The switch-off is logged in the public ledger. The handover fidelity is tested before the outgoing system shuts down — a cold-start drill that verifies the incoming steward can actually access and use the transferred records without manual reconstruction. The maps and evaluation results are gifted to the commons, available to any future emergency response in the region. The next response will start on day one with the knowledge that this one accumulated over six weeks.
The Kami departs. The river remains. The record helps the next steward start stronger.
What Could Go Wrong
Imperial creep is the risk that a capable agent, having succeeded in its assigned domain, begins extending into adjacent domains — first informally, then formally, until its original scope bounds are effectively meaningless. The fix is hard caps enforced by external infrastructure rather than self-reported, a non-expansion pact that requires affirmative fresh authority for scope changes, and capability monitoring by an independent oversight board that watches for the early signals of scope extension before they become entrenched.
Within-scope power accumulation is a subtler risk than imperial creep: a system can become dangerous without formally extending its scope if it accumulates leverage within its own lane. A sufficiently data-rich, sufficiently entrenched local system can effectively veto community decisions even if it never technically exceeds its mandate. The fix is the same as for imperial creep: caps verified by external infrastructure, not self-reported; independent resource audits that measure leverage as well as raw resource consumption; and capability monitoring that watches for accumulation of influence as well as accumulation of compute.
Treaty fragmentation is the risk that the proliferation of federation standards — each well-intentioned, each solving a slightly different problem — produces a landscape in which no two systems can actually interoperate because they speak different protocol dialects. The fix is a minimal core standard that covers the essential cases, adapters that translate between dialect variants, and conformance tests that verify real interoperability rather than nominal compliance. The standard should be polycentric — governed by a diverse coalition rather than any single actor — but it must be interoperable.
Zombie agents are systems that no one turns off because no one has clear authority to turn them off, or because everyone assumes someone else is responsible. The fix is sunset by default: a system that does not receive an explicit attestation of continued authority at regular intervals is presumed to be in its retirement phase and begins the succession process automatically. No attestation, no runtime.
Steward attachment is the most human of the risks: the builders and operators who have invested their professional identity in a particular system resist its retirement, treat succession planning as a threat rather than a governance requirement, and subtly or overtly obstruct the process. The fix is institutional rather than technical: term limits for named stewards, rotation of spokespeople so that no individual becomes publicly identified with the system's continuation, formal separation of the authority to operate a system from the personal brand of the person who built it, and a cultural norm that treats graceful succession as evidence of successful stewardship rather than evidence of failure.
Symbiosis and the 6-Pack
Chapter 8 does not replace the five chapters that precede it. It keeps them from becoming their own negation. Attentiveness without bounds becomes surveillance. Responsibility without bounds becomes a contractual trap. Competence without bounds produces indispensability. Responsiveness without bounds becomes a dependency relationship masquerading as care. Solidarity without bounds produces civic infrastructure that no one controls and everyone depends on.
The connection to Chapter 7 is particularly direct. The solidarity infrastructure that Chapter 7 establishes — selective-disclosure identities, portability standards, federated safety networks, protocol-level cooperation norms — is what makes symbiosis operationally feasible. Without portability standards, succession is a manual reconstruction rather than a verified handover. Without federation treaties, cooperation between bounded systems requires either a shared supervisor or repeated bilateral negotiation. Solidarity provides the civic infrastructure; symbiosis provides the discipline that prevents that infrastructure from being captured by any single actor who is tempted to treat it as proprietary.
And because the care practices that symbiosis enforces must embody democratic values all the way down — not merely produce good outcomes by coincidence — the treaties that solidarity establishes carry justice, equality, and freedom as operating constraints. A federation treaty that produces efficient coordination but excludes minority communities from its governance is not a solidarity instrument; it is a new form of capture wearing solidarity's clothing. Symbiosis insists that the form of the governance matters as much as the output.
A Culmination
This chapter closes the 6-Pack of Care by asking the question that should have been asked before any system was designed: what happens when it ends?
A system ends well when the community it served retains the capacity to go on without it — when the knowledge it accumulated has been transferred, the obligations it held have been inherited, and the next steward starts with the benefit of everything the first one learned. Designing for endings from the beginning is perhaps the most counter-cultural thing the 6-Pack asks of builders. The dominant incentives of the technology industry run in exactly the opposite direction: toward retention, accumulation, and the dependencies that make a system's continuation seem inevitable. The Kami model insists that this accumulation is not success; it is the early stage of a problem.
The discipline applies to the builders themselves. A stewardship that cannot survive its founders fading is not care — it is dependency. Genuine stewardship includes the willingness to step aside.
Imagine a river tended by local guardians — each keeping its bank, sharing warnings upstream and down, stepping aside when the season changes. The river does not need one ruler. It needs many stewards who know their stretch, and who know when to let go. That knowledge — of when to let go — is not a technical competence. It is a moral one. It is the knowledge that care is not ownership, that service is not possession, and that the highest expression of stewardship is to leave the thing you tended better than you found it, and to trust the next guardian to continue.
The Kami departs. The river remains.
Public measure: The headline public measure for Chapter 8 is exit readiness — whether the system can hand off, shut down, or retire without trapping the community inside it. Supporting diagnostics include portability drill success rates, handover fidelity scores, time-to-sunset from mandate expiry, diversity of overlapping stewards across a region, and whether service duties survive component retirements intact. These measures connect the symbiosis chapter back to the attentiveness of Chapter 3: retired agents that gift maps, evaluations, and operational records to the commons make the next system's first look better than the last system's first look was. The care loop continues, even when the components that ran it do not.