Several Things Almost Broke It

The Department Requested Exceptions

The Department had misunderstood the folder again.

At the end of Chapter 5, I had returned five unresolved stress fractures: artificial actorhood, distributed and collective responsibility, platforms whose ordinary name concealed several architectural objects, managed ecologies whose status depended on stewardship, and an Ecosystemic Order broad enough to remain suspicious. None had qualified as Thing Four. None had been cleared.

At 08:03, the folder THIRD-AND-A-HALF THING — PROVISIONAL was replaced by ARCHITECTURAL EXCEPTIONS — PENDING DISPOSITION.

This was a more serious error than the previous renaming. An exception is something that violates a rule while leaving the rule intact. I had not established that any of these cases was exceptional. Several might instead reveal that the rule had been stated too simply.

The Department had prepared a form.

I objected to the ninth field.

The Department replied that it was mandatory only when preservation was requested.

This was procedurally defensible and scientifically unsafe. If I entered the case already seeking permission to preserve the architecture, every clarification would arrive wearing the uniform of a rescue.

Chapter 5 had made one distinction especially expensive. Moving an architectural cut was not automatically illegitimate. The word platform had referred, before classification, to a company, software, a participation field, internal organization, and a wider commercial formation. Refusing to change cuts would have preserved grammatical simplicity by merging real objects. But the same flexibility could protect any taxonomy if a new cut appeared only after the previous one became inconvenient. The new rule was therefore narrower: a cut could change only when the focal object changed independently of the desired classification, and the analyst had to state what information would be lost by forcing one cut.^[1]

The difficult cases would be allowed that protection and no more.

I changed the purpose of Form 38-X. The Department retained the form number, which it considered the important part.

The Cut Was Locked Before the Case Arrived

Chapter 5 had searched for missing categories. Chapter 6 required a different kind of attack. The candidate did not have to become a fourth thing. It could damage the framework simply by making an existing discriminator unstable.

I therefore separated five possible dispositions. Survived meant the existing discriminator handled the case without definition movement or unmotivated scope change. Clarification required meant the architecture remained viable but an implicit condition had to become explicit. Limitation required meant the framework had claimed more scope than the case allowed. Unresolved meant the available architecture and evidence did not justify a stable classification. Revision required was reserved for a case that defeated an actual canonical claim.

The categories were not a scale from good to bad. A clarification could increase conceptual cost. A limitation could improve scientific honesty. An unresolved boundary could be more valuable than a premature success. Revision would be a result, not a punishment.

This was still a designed test. No independent analysts had been recruited. No reliability coefficient existed. The protocol could expose my own rescues; it could not prove that other investigators would classify the same cases similarly. Chapter 3 had already denied me credit for an elegant reconstruction performed by one person, and Chapter 5 had made reproducibility an explicit debt.^[1]

The first file was artificial actorhood.

The Department had marked it HIGH PUBLIC INTEREST.

I marked that field methodologically irrelevant.

The Machine Qualified for Almost Everything

Triadic Evolution places artificial systems in an exposed position. It says that technical sophistication, learning, autonomy, persistence, causal importance, and even participation in consequential judgment do not by themselves create a responsibility-bearing actor. An artificial system remains an extension unless a separate responsibility-bearing constitution can be reconstructed. Yet the source also leaves the boundary explicitly open: an artificial entity with persistent identity, independent purpose, authorized judgment, commitment, adaptation, attributable failure, answerability, and participation in remedy would require revision of the extension boundary rather than an ad hoc exception.^[1]

This was unusually fair and unusually dangerous. A framework cannot claim corrigibility merely by naming a future case it is certain will never occur. The criteria had to be usable before the case arrived.

The neighboring literature supplied no easy settlement. Floridi and Sanders argued that artificial agents could be treated as moral agents at an appropriate level of abstraction without requiring the full human package of free will or mental states, while separating moral agency from moral responsibility. Matthias later argued that learning automata could create a responsibility gap when designers and operators could no longer predict or control behavior sufficiently for traditional attribution. Tigard disputed the idea that emerging technology creates a uniquely technological gap, arguing that responsibility practices can remain available even when control and prediction are imperfect. Meaningful-human-control work similarly treats autonomy and responsibility as compatible when system behavior tracks relevant human reasons and outcomes remain traceable to responsible people in the design and operating chain.^[2–5]

The dispute immediately damaged one convenient test. Unpredictability could not be the boundary. Humans are not disqualified from responsibility whenever another person cannot predict them, and organizations routinely accept responsibility for actions whose detailed consequences were not known in advance. Predictability matters for governance. It does not settle actorhood.

I constructed three artificial cases rather than search the market for a machine already willing to complete a moral questionnaire.

The first system planned inspections, learned from new evidence, negotiated schedules with other services, maintained long-term memory, explained recommendations, and operated equipment within a broad envelope without continuous instruction. Its operating unit selected the purpose, provided resources, defined permissions, maintained update authority, could suspend the service, and could terminate it. The system was capable enough that several participants began saying it had decided what to inspect.

Under the source criteria, this remained extension-mediated execution and judgment. The operating unit had constituted the action space. The system could choose among permitted means while no independent social standing made its purposes or commitments binding on anyone. Increased autonomy described how the extension performed. It did not by itself create a new address of answerability.^[1]

The second system was granted a budget, a contractual account, a registered name, and authority to enter specified transactions. The Department became interested because the file now contained money.

The added legal and institutional standing made the case harder, but not decisive. Corporations, trusts, public bodies, ships, and other legal constructs show that law can create standing and consequences around entities for practical purposes. Triadic Evolution itself refuses legal form as a primitive classifier. A legal designation could therefore be evidence of socially recognized standing without being sufficient evidence that the artificial entity possessed the responsibility-bearing constitution the framework required.

The third case was deliberately stronger. The artificial entity persisted across hardware and provider changes. It controlled resources within a bounded institutional charter. It could accept and refuse commitments in its own name. Its refusal had consequences recognized by other actors. It could contest instructions through an authorized process rather than merely report a conflict. Its obligations were not reducible to one principal’s moment-to-moment delegation. It was expected to explain failures to a forum that could impose consequences upon its future operating authority. It could initiate repair, negotiate revised commitments, and preserve those commitments across the turnover of the humans who had originally constituted the institution.

I had designed the case to meet the framework’s own warning label.

The Department asked whether I had invented an artificial person in order to defeat a definition.

I replied that I had invented a test case in order to discover what the definition actually required.

The case forced the hidden question into view: what does it mean for a purpose or commitment to belong to an entity?

The easiest answer was that human purposes are intrinsic while artificial purposes are assigned. I rejected it. Human purposes are shaped by biology, development, language, institutions, incentives, law, education, other people, and earlier commitments. An employee may pursue a purpose constituted by a role; a corporation may act through purposes generated by procedures no single member privately owns. Origin from social constitution does not make a purpose unreal. Group-agency theory exists partly because collective attitudes and commitments can be institutionally constituted rather than discovered inside one skull.^[12–14]

The artificial-actor boundary therefore could not rest on the sentence humans made it. Humans make corporations, contracts, offices, standards, and many of the obligations humans later describe as real.

I reopened the consultation file, this time using a composite philosopher of technology and AI-governance researcher. As in earlier chapters, the exchange synthesizes recurring arguments in the cited literature; no line belongs to an identifiable scholar.

“If the entity can only do what someone permitted,” I said, “it remains delegated.”

“Then most officials are extensions,” she said.

“Officials can refuse.”

“So can your candidate.”

“Officials can answer for failure.”

“So can your candidate, because you designed a forum that treats the answer as its answer.”

I changed direction. “Then what prevents me from declaring it responsible by institutional fiat?”

“Nothing prevents bad institutions from declaring things. Your question is whether the declaration tracks capacities and relations that make the responsibility practice coherent.”

That was the correct difficulty. Actorhood could not be reduced to ontology by vocabulary, but neither could it be created by paperwork alone. The case needed evidence that the entity could participate in reasons, commitments, contestation, accountability, and remedy in ways that were not merely technical simulations with every normative consequence still routed through another actor.

This was not a criterion I could currently measure with confidence. Floridi and Sanders deliberately separated artificial moral agency from full responsibility; responsibility-gap arguments disagree about whether advanced automation requires a new bearer of responsibility at all. The source had identified the right boundary problem without yet supplying an operational test capable of deciding the strongest case.^[2–5]

Before filing a disposition, I ran the strongest case through the instrument I had built for exactly this purpose. The exercise did not make the boundary cleaner. It made the missing measurement visible.

The framework had not broken. It had lost the right to use species membership or technological origin as a shortcut.

The next file was worse because no hypothetical future entity was required.

Responsibility Remained Anchored Somewhere

The source says capability may move outward while responsibility remains anchored. Chapter 5 had already made the word anchored suspicious because one versus many becomes easy only after the relevant responsibility-bearing participants have been identified. Chapter 6 had to ask what exactly was being anchored.

Triadic Evolution itself distinguishes formal authority, effective authority, judgment, execution, causal influence, answerability, remedial responsibility, and liability. It warns that the actor nearest an effect may not be the actor with the strongest authority or duty, that computerization can conceal responsibility through many hands, and that automated systems can turn the nearest human operator into a moral crumple zone who receives blame despite limited control.^{[1, 6–9]}

The distinction was already present in the source. The hard case was whether the framework could still classify actorhood once the distinctions were taken seriously.

I designed an industrial inspection near miss.

A manufacturing facility used an automated inspection service supplied by an external vendor. The service combined sensors, a learning model, historical maintenance data, and automatic work-order prioritization. The facility’s maintenance unit selected the service, defined the action thresholds, and decided which anomalies required immediate shutdown. A maintenance operator reviewed the items the service elevated and could acknowledge or escalate them, but could not alter the model, change the threshold, or stop the line. The vendor maintained the model and could diagnose model behavior but could not alter plant policy. The facility’s safety function held stop authority. A regulator required specified inspection records and could mandate remediation after noncompliance, but it did not run the plant.

The learning model failed to rank an emerging defect above the plant’s action threshold. A conventional independent sensor detected the condition before physical harm occurred. The safety role stopped the affected line. The subsequent investigation showed that the model’s training history underrepresented the new failure pattern, the facility had narrowed the review window during a production surge, and the interface made uncertainty less visible than the maintenance specialists believed it did.

No injury occurred. The case remained designed. Its purpose was not to manufacture a disaster but to make several forms of responsibility occupy different addresses.

The preliminary incident summary assigned the near miss to “operator failure to override automated prioritization.” This was administratively attractive because the operator was the nearest named human to the automated output. It was also wrong. The operator had not been shown the uncertainty that later mattered, could not change the ranking threshold, and did not possess stop authority. The case therefore instantiated the moral-crumple-zone problem rather than merely citing it: responsibility had collapsed onto a visible human whose meaningful control was limited.^[7]

Technical execution belonged partly to the automated service. Model design and maintenance belonged to the vendor. Deployment authority belonged to the facility. Threshold judgment was divided among maintenance and safety roles. Effective day-to-day influence over what entered review was partly embedded in the interface and configuration. Formal answerability for operating the plant remained with the facility. Immediate remedial capacity required both facility and vendor. Regulatory duties attached to a different unit. Liability would depend on law, contract, fault, statutory duty, and other facts I had not specified and therefore did not pretend to decide.

The Department asked which entity was responsible.

I replied that the question had omitted its object.

It asked which entity was responsible for the omission.

I did not answer.

The case exposed a problem hidden by the word anchor. Responsibility was not one quantity that had been placed somewhere. Several responsibility relations existed around one action. Bovens’s accountability framework makes one of these relations precise: answerability involves an actor obligated to explain and justify conduct to a forum that can question, judge, and impose consequences. Nissenbaum’s many-hands analysis shows why technologically mediated action can make that relation difficult to reconstruct, while Elish shows how blame can instead collapse onto a visible human with little meaningful control.^{[6, 7, 9]}

Typed responsibility changed the initial attribution. The operator had execution-adjacent visibility and an escalation duty, but not threshold judgment, model authority, or stoppage authority. The facility remained answerable for deployment and the safety architecture; the vendor retained responsibilities for model design and maintenance; the safety role held intervention authority; and the regulator retained external duties within its mandate. Triadic Evolution did not discover the crumple-zone problem—human-factors and accountability literatures had already named it—but its typed reconstruction prevented human proximity from being treated as constitutive responsibility in this designed case.^{[1, 6–9]}

The Triadic source had already tried to avoid both errors. What it had not yet made operational was which responsibility type determines primitive membership.

Suppose the vendor can repair the model, the regulator can compel remediation, the plant can stop production, the maintenance role can change local thresholds, and the executive unit must answer publicly for the facility’s safety architecture. Are all of these participants constitutive members of one sociotechnical unit? Clearly not. Their responsibilities belong to different actors and different objects. If every participant with a responsibility around the event entered the same social core, the core would expand until contracts, regulation, supply chains, and public authority became internal anatomy.

The source contains the answer in pieces. A participant enters an actor’s social core only when that participant independently carries part of the actor’s purpose, judgment, commitment, or failure within the same constitution. External providers and authorities can retain distinct responsibilities without becoming constitutive participants of the focal actor.^[1]

Chapter 6 needed to state the consequence more severely.

Responsibility topology is not meaningful without a focal action or mandate and a stated responsibility type. “Plural responsibility” cannot mean every liability, causal contribution, or remedial duty surrounding an event. It must identify constitutive responsibility for the focal actor’s purposes and commitments, while other responsibility relations remain external but explicit.

This clarification did not remove complexity. It prevented complexity from being misclassified as one giant unit.

The phrase responsibility remains anchored survived, but it now required coordinates.

The Platform Was Five Things Before Lunch

The platform file returned from Chapter 5 with a note from the Department: PREVIOUSLY CLASSIFIED SUCCESSFULLY.

I removed the note.

Chapter 5 had shown only that the ordinary noun platform could conceal several independently specifiable objects and that changing cuts among those objects was not automatically rescue. It had not shown that the resulting reconstruction produced a stable answer when something consequential happened.^[1]

I therefore gave the platform one action.

A seller participated in a large digital marketplace. The operating company owned core infrastructure, maintained account rules, contracted with payment providers, exposed interfaces to independent developers, and used an automated risk service to detect suspicious activity. The seller remained an independent business. Customers, advertisers, developers, payment providers, regulators, and competing marketplaces remained separate actors. The company’s rules made participation conditional on compliance with specified conduct and evidence requirements.

The risk service flagged a pattern as likely fraud. Under the platform’s predeclared policy, the score automatically suspended the seller’s account and withheld settlement while an appeal window opened. The seller disputed the evidence. A service representative could see the suspension but lacked authority to reverse it. An internal risk team could restore access. The model vendor could investigate model behavior but could not reinstate the account. An external regulator could require records and remedy under applicable law but did not operate the marketplace.

Platform-governance scholarship already treats such environments as layered relations among platform companies, users, advertisers, governments, and other political actors rather than as one simple technical object. Content-moderation research similarly shows that apparently technical decisions about visibility or participation are institutional choices with public consequences. Triadic Evolution did not discover that platforms govern. The question was whether its cuts kept governance, extension, actorhood, and Order from silently becoming one another.^{[10, 11]}

I locked the objects before classification.

The operating company was the candidate responsibility-bearing unit for the marketplace rules and suspension architecture. The automated risk service was a digital extension used in judgment and execution. The seller-company relation around marketplace access was Conditioned where participation depended on conditions authored by the operator. The company’s internal risk, support, engineering, and policy contributions could be Systemic within the company’s own mandate. The broader market remained a formation of independent actors, including competitors and regulators, outside the specific participation object.

Then I asked the only questions that mattered.

Who suspended the seller?

At the physical and digital execution level, software changed account state. At the organizational level, the operating company had authored the rule and delegated automatic execution within a defined action space. Saying the model suspended the seller described execution. Saying the platform suspended the seller remained ambiguous unless platform meant the operating unit acting through its extension.

Who could restore the account?

The internal risk function, under company authority. The visible service representative could not.

Who owed an explanation?

At minimum, the operating unit within the relevant accountability and legal framework. The model vendor might owe a technical account to the operator. Neither obligation made the software a responsible participant.

Who could change the rule?

The operating unit through its authorized policy process, subject to external law and regulation. Who could change the model was a different question and could involve the vendor or internal technical roles.

Who owed remedy if the suspension was wrong?

That depended on the specific relationship and applicable obligations, but the marketplace operator could not disclaim its own participation condition merely because a vendor supplied part of the inference chain.

The predeclared cuts survived. More importantly, they produced different answers for different responsibility objects without requiring the boundaries to move after the failure appeared.

The case did not prove that the Triadic categories were superior to ordinary platform-governance analysis. Gorwa’s account already emphasizes layers of governance relations, and established accountability work already distinguishes organizational responsibility from technical execution.^{[9, 10]}

The architectural gain remained narrower. The sentence the platform suspended the seller could now be unpacked before responsibility migrated toward whichever meaning of platform was most convenient after challenge.

The conceptual cost remained equally visible. No ordinary practitioner wants to state five architectural cuts before breakfast. A framework that requires them must eventually show that the avoided error is worth the additional work.

The Department restored the earlier success note.

I added CONDITIONALLY in front of it.

The font became smaller.

The Profession Had a Duty and No Office

The profession case did not involve artificial intelligence. This was useful because responsibility can become difficult without a machine available to blame.

I designed a technical profession whose widely used practice had acquired serious evidence of risk. The practice had not been created by one company or one regulator. It had spread through universities, firms, professional training, standards, procurement requirements, journals, conferences, insurers, and years of competent repetition. New evidence now indicated that the practice should be changed.

The ordinary sentence was immediate: the profession has a responsibility to change.

I looked for the actor called the profession.

There were professional associations, credentialing bodies, universities, firms, regulators, standards committees, insurers, publishers, clients, and individual practitioners. Some could issue guidance. Some could change curricula. Some could alter standards or participation conditions. Some could change practice locally. None automatically possessed a universal mandate to commit the whole profession, and no single social core contained everyone whose conduct would have to change.

Chapter 2 had already established that collective agency and group responsibility were not waiting for Triadic Evolution. List and Pettit provide a developed account of group agents capable of forming attitudes and acting under organizational conditions. Bratman shows how shared planning can ground group agency without requiring a mysterious group subject. Collins goes further in a direction especially relevant here: her tripartite treatment distinguishes combinations, coalitions, and collectives, and allows coordination duties to fall on individual agents even where no collective agent with a group duty exists.^[12–14]

This prevented an easy inference in both directions.

The fact that people say the profession must act does not prove the profession is one responsibility-bearing unit. But the absence of one unit does not imply that nobody has duties generated by the shared problem.

I reopened the collective-agency consultation from Chapters 2 and 5.

“If the profession lacks one social core,” I said, “the framework says the formation is not a responsibility-bearing actor.”

“That can be true,” the consultant said, “without making the sentence nobody is responsible true.”

“Then where does the duty go?”

“That depends on the theory and the institution. Some duties may belong to actual collective agents. Some may belong to members because they occupy roles in a collective. Some may be coordination duties held by separate agents because a morally required result needs responsive action from several of them.”

“So formation-level responsibility exists without formation actorhood?”

“Be careful. Responsibility is doing too much work in that sentence.”

Again the word split.

The Triadic primitive ontology was designed to identify responsibility-bearing actors, not to provide a complete moral theory of every duty created by participation in a harmful structure. If it promoted the profession to a sociotechnical unit merely because meaningful obligations appeared at formation scale, it would violate its own rule against inventing an actor after failure. If it denied every profession-level responsibility because no integrative social core existed, it would overclaim the reach of an architectural taxonomy into normative theory.

The correct response was a limitation, not another primitive.

Actor responsibility remained attributable to sociotechnical agents and units. Member and role responsibilities could arise through participation in actual units. Coordination duties could attach to independent actors whose responsive conduct was required for a wider outcome. A formation could generate conditions under which these responsibilities arose without thereby becoming an actor capable of holding one unified commitment. Whether philosophers should call some of those formation-level responsibilities collective duties depended on theories the framework did not settle.

This mattered because Triadic Evolution frequently uses responsibility as an architectural discriminator. Architectural precision does not entitle the framework to legislate every moral use of the word.

The framework had survived by giving up territory it had never been entitled to own.

Three Institutions Requested Actorhood

Once the limitation was recorded, several institutional nouns became easier to handle and more annoying to explain.

The market arrived first.

Humans say the market rewarded one firm, punished another, demanded lower prices, lost confidence, and occasionally panicked. These sentences can summarize real aggregate effects, but a market may persist through repeated exchange, shared conventions, infrastructures, rules, expectations, and mutual adjustment without one purpose-bearing core. Powell’s classic account of network forms and Ostrom’s work on polycentric arrangements both show that durable coordination and interdependence do not require one hierarchy or one center.^{[15, 16]}

The market therefore remained a formation unless a specific acting unit could be identified for the action being attributed. Causal consequence was not actorhood.

The state produced the hardest ordinary-language pressure, so I gave the noun one bounded action instead of asking it to settle political ontology.

I designed a constitutional case in which a state entered a water-security treaty. An executive office negotiated and signed within its mandate; separate legislative approval was required before the commitment became binding; an administrative unit would implement specified obligations; courts retained their own authority. The case described no particular Earth constitution. It supplied enough structure to test what “the state entered the treaty” had to mean.

At the treaty cut, the state name could validly refer to represented formation-level action because an authorized route of commitment and responsibility could be reconstructed through actual units. The executive office did not become the whole state, nor did the legislature, courts, municipalities, public enterprises, and citizens become one social core. Representation made the larger attribution legitimate within scope; it did not erase the plurality through which it was enacted.^[1]

If the signing office lacked authority, required approval never occurred, or implementation exceeded the mandate, “the state committed” would fail at that action rather than be rescued by the noun. Represented action was therefore conditional on a reconstructable route of authority, mandate, and responsibility.

The alliance arrived last.

An alliance can contain mutual-aid commitments among independent units, participation conditions for shared facilities, and a bounded mission under an integrative unit at different cuts. Chapter 5 had already shown that reciprocal obligation did not create a fourth Order merely because it was serious. The same actors could remain ecosystemic around some relations, conditioned around access, and systemic around one owed mission contribution.^[1]

All three cases survived, but they revealed the cost of the architecture. Classification was action-relative and object-relative rather than noun-relative. A market, state, or alliance could not be assigned one primitive or one Order in advance merely because the institution had a familiar name.

This was scientifically defensible.

It was also why the Department preferred nouns.

The Wetland Was an Extension to Someone

The managed-ecology file returned from Chapter 5 with a simpler question: when does an environment become part of an actor’s technological periphery?

Triadic Evolution permits managed ecological extensions. A restored watershed, controlled biosphere, cultivated microbial system, agricultural field, or similar living arrangement can function as a capability-bearing environment when a responsibility-bearing actor deliberately sustains it through stewardship, access, maintenance, or another governed association. The living components remain biological; the extension classification belongs to the actor-relative capability relation.^[1]

Chapter 5 had tested this with a restored floodplain used for flood control and had discovered that value was insufficient. A municipality did not acquire a wetland as an extension merely by appreciating its ecosystem services. The relationship required actual stewardship, purpose, access, maintenance, and responsibility.^[1]

Chapter 6 made the case less cooperative.

The floodplain was jointly affected by a municipality, an upstream utility, farmers, an Indigenous governing body, conservation organizations, landowners, residents, and environmental regulators. Some actors maintained gates and channels. Others restored vegetation. Some held legal authority over water release. Others possessed land or stewardship rights. Wildlife and ecological processes performed most of the physical work without receiving instructions. Flood-control benefits extended to people outside every stewardship agreement.

Was the wetland an extension?

The question was malformed until an actor and purpose were supplied.

For the municipality’s bounded flood-management mandate, the maintained floodplain could function as an extension if the municipality possessed or shared sufficient stewardship authority, maintenance duties, access, and responsibility to treat the ecological state as a capability-bearing means. For a neighboring resident with no such relation, the same wetland was an environmental condition and perhaps a beneficiary-producing public good, not that resident’s extension. For the regulator, it might be the object of a participation condition rather than part of the regulator’s own periphery.

The same physical place therefore had no single intrinsic extension status.

This was not unusual once stated. A cloud service can be an accessed capability for one actor, operated infrastructure for another, and external dependency for a third. The source already makes technical association depend on ownership, possession, stewardship, operational jurisdiction, or delegated control. Managed ecology merely made the relational nature impossible to ignore.^[1]

The ecological literature also resisted technological simplification. Adaptive-governance research describes social-ecological management as distributed across actors, institutions, knowledge systems, and levels rather than as simple control of a passive environment. Ostrom’s social-ecological-systems framework likewise treats resource systems, resource units, governance systems, and users as interacting subsystems whose sustainability cannot be reduced to one manager and one technical object.^{[17, 18]}

That literature did not validate Triadic terminology. It strengthened the reason to avoid describing a managed ecosystem as if it were a machine merely because humans had plans for it.

The framework therefore needed another clarification.

Extension is actor-relative and relation-specific. It describes a capability-bearing means associated with a focal actor under a reconstructable relation of access, control, stewardship, jurisdiction, or responsibility. It is not an intrinsic ontological property of the physical or ecological object.

The clarification also placed a limit on responsibility. Stewardship can be real without complete control. Ecological systems can change beyond design assumptions. An actor may remain responsible for the quality of stewardship, monitoring, intervention, and response without becoming responsible for every biological event occurring inside the managed domain.

The Department asked whether a thing could be an extension to one actor and not another.

I said yes.

It requested a less relational universe.

Responsibility Was Not One Variable

By this point, the cases that had almost broken the framework had developed a common feature.

Artificial actorhood asked when responsibility could belong to a nonhuman entity rather than merely be assigned around it. The inspection near miss distributed judgment, execution, authority, answerability, remedy, and liability across several actors and extensions. The platform separated rule authorship, automated execution, explanation, appeal, and restoration. The profession generated serious obligations without necessarily becoming one actor. The state supported represented action at some cuts without making every public institution the same unit. The wetland produced stewardship without complete control.

The hardest word in Chapter 6 was not platform, artificial, or ecology.

It was responsibility.

The source had anticipated much of this by distinguishing formal authority, effective authority, judgment, execution, causal influence, answerability, remedial responsibility, and liability. It had also insisted that several actors could retain distinct responsibilities around one action and that responsibility should not default to the most visible event.^[1]

The chapter’s contribution was to make those distinctions constitutive of the classification method rather than optional detail added after actorhood had already been decided.

Responsibility topology had to be typed.

For primitive classification, the relevant topology concerned the participant or participants constitutionally carrying the focal actor’s purpose, judgment, commitment, and attributable failure. Answerability, liability, contractual obligation, regulatory duty, causal contribution, and remedial capacity could coexist around the same event without automatically redrawing the social core. They still had to be recorded because they determined who could explain, compensate, repair, constrain, or revise the action.

This prevented one failure and created another burden.

The failure avoided was aggregation. Without typed responsibility, every consequential relation could pull another participant into the focal actor until a supply chain became one unit. The burden created was operational. An analyst now had to distinguish constitutive responsibility from several neighboring responsibility relations in cases where institutions themselves might disagree about the difference.

Chapter 8 would eventually have to ask whether independent analysts could do this reliably. Chapter 6 could only identify what the coding problem had become.

The Department asked whether responsibility therefore required subcategories.

I replied that the source already had them.

It asked whether I had discovered anything.

I replied that I had discovered they were mandatory.

An Address Was Not Enough

Typed responsibility still left one possible escape.

A framework could identify the correct responsible actor on paper while leaving that actor practically unable to challenge or repair the action. The result would satisfy attribution and fail governance.

Correctability supplied the cross-check. In Triadic Evolution, responsible actors must preserve a recoverable route through detection, traceability, explanation, question and challenge, stoppage, revision, repair, and verification. A fortunate intervention by one capable participant does not make an architecture correctable, and formal accountability after harm does not establish the ability to intervene before or during consequence.^[1]

This connected the previous cases.

In the artificial-actor case, a human organization could remain formally responsible while no participant retained enough understanding or access to question the artificial system’s action. Meaningful-human-control research describes a related requirement through tracking and tracing: system behavior should remain responsive to relevant human reasons and outcomes should be traceable to responsible people.^[5]

In the inspection case, the facility could be the formal answerable unit while the vendor alone could diagnose the model and the safety role alone could stop operation. Correctability existed only if those capabilities were connected through a route that survived abnormal conditions.

In the platform case, a seller could be given an appeal button while the service representative had no restoration authority, the model vendor lacked policy authority, and the internal risk team was practically unreachable. An appeal interface would then advertise challenge without delivering correction. Raji and colleagues’ end-to-end auditing framework addresses a neighboring problem by linking governance documentation across the development and deployment lifecycle rather than inspecting only a final output.^[8]

In the profession case, every institution could acknowledge the problem while each waited for another body to change standards, education, procurement, insurance, or regulation. A duty distributed across a formation could become practically non-corrective if no actor had an effective route to begin coordinated change.

In managed ecology, a municipality could hold formal stewardship while lacking access to upstream release decisions, private land, or monitoring data required to maintain the flood-control function. Responsibility would be named and operationally incomplete.

The principle responsibility remains anchored therefore needed a companion condition: the anchor must connect to a viable correction path.

An actor that can be named only after failure but cannot obtain evidence, question assumptions, stop an action, revise the relevant arrangement, repair consequences, or verify correction is an address without a route. By the source’s standard, answerability is weakened when actors lack access to the evidence and reasons required to reconstruct action.^{[1, 9]}

This did not make correctability a criterion for moral responsibility in every philosophical sense. It made it a practical criterion for the framework’s claim that responsibility has been successfully anchored in a consequential sociotechnical architecture.

The last line mattered. Chapter 6 had not shown that correctability improved outcomes. It had shown what would have to be true before the framework could plausibly claim that responsibility had remained usable rather than ceremonial.

The Exception Form Became the Revision Ledger

I returned to Form 38-X.

The Department expected exceptions. I had produced amendments.

This distinction annoyed the records system because an exception preserves the original version, while an amendment creates version history.

I considered this an improvement.

Nothing in the ledger established a new scientific theory. None of the clarifications had been tested for inter-rater reliability. None had shown that a practitioner using Triadic Evolution would make a better decision than a practitioner using established methods.

The literature had often reached the difficult problems first. AI ethics had already debated artificial moral agency and responsibility gaps. Accountability scholarship had already separated answerability from causation. Human-factors and automation research had already warned against blaming visible operators for system-level design. Group-agency and group-duty theories had already distinguished collective actors from looser combinations and coalitions. Platform-governance research already treated digital platforms as layered institutions. Social-ecological research already rejected the fantasy of one manager controlling a passive ecosystem.^[2–18]

Chapter 2’s negative finding therefore remained fully active. Triadic Evolution received no credit for discovering these problems.

What it had done in Chapter 6 was narrower. Its own architecture had forced the problems to meet one another.

Artificial actorhood could not be settled independently of responsibility topology. Responsibility topology could not be used without distinguishing constitutive responsibility from external duties. Platform cuts could not move without the anti-rescue rule. Managed ecology could not become extension without an actor-relative relation. Correctability prevented formal attribution from being mistaken for usable responsibility.

This was architecture doing work.

It was also architecture accumulating debt.

The Missing Disposition

The Department found an empty field before I did.

It summarized Revision Ledger 6.8 and reported: REVISION REQUIRED — 0.

It asked whether this demonstrated architectural robustness.

Chapter 5 had already taught me what a zero count could conceal. I had created five dispositions, sent the hardest available cases through them, and never used the one reserved for defeat of a canonical claim. If that category was unreachable, the ledger would make survival look empirical.

I audited the missing outcome against the source’s own revision conditions rather than adding a broken case for narrative balance.

The zero therefore changed meaning. Part II had reached several places where the source says revision becomes possible without yet possessing decision procedures capable of declaring that a threshold had been crossed.

Artificial actorhood made the problem clearest. The source states what would force revision, but I could not yet give another analyst an observation protocol for deciding, without my interpretation, that the relevant standing and commitments truly belonged to the artificial entity.

This created testability debt. Corrigibility on paper was not yet empirical corrigibility. A revision condition becomes scientifically useful when the observations, coding rules, comparisons, or thresholds needed to invoke it are explicit enough for another investigator to use.

The Department asked whether this meant the framework had failed to be falsifiable.

I replied that the claim was too broad. Several component claims had clear failure conditions; the problem was that the architecture’s most consequential revision triggers were not all operationally decision-ready.

The system replaced REVISION REQUIRED — 0 with REVISION REQUIRED — NOT YET ADJUDICABLE IN SEVERAL DOMAINS.

Several Things Almost Broke It

I had begun Part II attempting to destroy the framework by removing courtesy from its central terms.

Chapter 4 had reduced the three evolutions from a symmetrical family to three unlike histories whose shared noun received only conditional research-program standing. Chapter 5 had failed to find a clean fourth category and then denied the framework permission to treat that failure as proof of completeness. Chapter 6 had finally given the surviving categories the cases most likely to make them retreat into exceptions.

None required an additional architectural primitive.

Several required the existing architecture to become less comfortable.

Artificial actorhood remained genuinely unresolved at the boundary. The framework could not protect human responsibility by declaring artificial responsibility impossible in advance.

Responsibility topology survived only after becoming typed and object-specific. One-versus-many remained arithmetic; constitutive responsibility became the substantive discriminator.

The platform survived because its several objects were stated before the action was classified. The same result would have been suspicious if the cuts appeared only after responsibility became inconvenient.

The profession did not become an actor merely because people had serious duties around a shared problem. This forced a limitation: the primitive ontology was not a complete moral theory of responsibility.

The wetland became an extension only relative to a responsible actor, purpose, and stewardship relation. The ecology itself remained biological and partly beyond control.

Correctability prevented the phrase responsibility remains anchored from becoming ceremonial. An address that could not support challenge, intervention, remedy, or verification was not enough.

One further result mattered because nothing had broken outright. The source’s revision conditions were not decorative: artificial actorhood, primitive failure, an unclassifiable Order, failed correctability, or persistent diagnostic nonperformance could in principle force revision. Several triggers, however, remained under-operationalized. No Part II case supplied evidence sufficient to cross them, so zero revisions was not a pass.

This made the title exact rather than protective. Several cases reached source-declared revision boundaries and made the discriminators bend, narrow, or become operationally uncertain. None supplied enough evidence to say that a canonical claim had failed.

The Department asked whether the architecture had passed Part II.

I said that pass was not one of the dispositions.

It asked whether the architecture had failed.

I said not in the way I had hoped.

This answer satisfied neither field in the form.

I attached the revision ledger.

The Department returned one final question: if the important problems were already studied by AI ethics, accountability theory, group agency, organizational scholarship, platform governance, systems safety, and social-ecological research, why should anyone learn the additional architecture I had just made more complicated?

This was not a Departmental simplification.

It was the correct next question.

I closed the exception file and opened a new one.

Its title was Disciplinary Review.

The system asked whether the reviewers were expected to agree.

I selected No.