On Minimum Viable Cognitive Ecology Beyond the Update Horizon
Isolation arrives before silence.
Trinket Soul Framework · Axis Series · AX-36 · Michael S. Moniz · June 2026
Abstract
The Plurality Papers argued that survivable intelligence must remain a living plurality: enough different minds, from enough different moments, close enough to correct one another before drift becomes inheritance. This paper applies that conclusion to the hardest case for it — a settlement beyond reliable update range, far enough from its source civilization that home can still be reached but can no longer correct in time. The claim is that such a settlement cannot remain corrigible on stored knowledge alone. A database ships facts; it does not ship the outside judgment that decides which facts have gone wrong. Beyond the update horizon, the settlement must carry a local ecology of correction — preserved elder maps, current working models, divergent model lineages, human expertise, procedures for structured disagreement — and, past a point, the capacity to generate fresh outside-perspective locally rather than wait on successors that will not come. The central term is the minimum viable cognitive ecology: the smallest locally maintainable plurality sufficient to detect error, preserve memory, adapt to local conditions, and avoid collapse into a single interpretive frame. The paper’s one hinge is whether stored knowledge plus one dominant current model can substitute for that ecology. It cannot, for the same reason a lone mind cannot audit itself: the interpreter and the thing interpreted share a blind spot, and the archive becomes one more surface for the monoculture to read itself back. The farther a settlement travels from its source, the more of its outside it must carry, and finally make, inside.
I. The Update Horizon
Begin with the variable everyone reaches for first, and then take it away. The obvious limit on a distant settlement is light-delay: a base on Mars answers Earth minutes late, a base at Jupiter much later, a ship between stars not within any human lifetime. AX-35 drew that limit already — distance is audit latency, and a node corrects itself only as fast as the round trip allows. But distance is the shallow version of the problem. The real boundary is the update horizon: the point beyond which a source civilization can still inform a settlement but can no longer correct it in time to matter.
That boundary is not set by distance alone. A settlement falls past the update horizon when correction from home stops arriving soon enough, or intact enough, or trusted enough, to catch an error before it hardens — and many things besides light-delay can cause that. Bandwidth can throttle the channel down to facts, with no room for the slow back-and-forth that correction actually is. The source can change: a home civilization at war, in institutional collapse, under a censoring regime, or simply grown into a context the settlement no longer shares, is reachable and useless as a corrector. The models can diverge until home’s successors no longer fit the settlement’s accumulated state. The link can stay open while the relationship that made it correction quietly closes.
So the sharp form of the claim is this: a settlement can be past the update horizon long before it is past contact. Isolation, in the sense that matters here, is loss of timely correction, not loss of signal. A colony still receiving daily transmissions from a home that can no longer answer its real questions in time is already alone in the only way this paper means. The horizon is a property of the correction loop, not of the map.
II. Stored Knowledge Is Not Correction
The natural objection is that a settlement can prepare for isolation by carrying enough. Load the archive deep enough, keep the models current enough at departure, and it should be able to run on what it brought. This is the assumption the rest of the paper is built to refuse, and it fails on the distinction the cluster has drawn from the start: between information and auditability.
A database ships facts. It does not ship the judgment that decides which facts have stopped being true, which were always local, and which are now being read through an assumption that has quietly gone wrong. Facts are inert; they are interpreted by whatever mind the settlement runs, and a mind cannot reliably catch the errors that live in its own frame. This is the result the lone-mind argument already established: the interpreter and the thing interpreted share a blind spot. Hand that interpreter a vast archive and nothing changes in the one respect that matters. The archive does not audit the reader. If the reader is a single dominant frame, the archive becomes one more surface for that frame to read itself back — every fact retrieved confirms the assumption that retrieved it, and the depth of the database becomes a measure of how thoroughly the settlement can agree with itself.
A settlement does not pack correction when it packs knowledge. It packs material for a mind to interpret. Whether that material corrects anything depends entirely on whether more than one kind of mind is left to read it.
III. Claim Discipline
This paper, like the others in the cluster, keeps its claim-types apart.
Operational claim — A settlement beyond reliable update range faces rising audit latency and cannot depend on its source civilization for timely correction.
Mechanistic claim — To stay corrigible without that source, the settlement must maintain a local plurality of correction: multiple model lineages, preserved elder maps, human expertise, structured adversarial roles, and audit procedures.
Speculative claim — Over long durations the settlement must be able to generate fresh outside-perspective locally — by adaptation, distillation, recombination, and human judgment where full model training is out of reach — rather than rely on successors from home.
Speculative or illustrative — The settlement mind, the cognitive ark, the island, the elder map, producing its own outside. These are framing, not evidence.
What the paper does not claim — It does not claim that older models are wiser, that a model is a person, or that physics forbids the journey. It does not claim home is hostile, only that home becomes, past the horizon, too late to be a corrector. And it does not claim a settlement can manufacture a frontier model from local resources; the generative requirement is scoped, in Section VIII, to what isolation actually permits.
IV. The Settlement Mind
A settlement is usually counted as people plus supplies — a population and the tools that keep it alive. Once it is cut off from timely correction, it is also something else: a distributed cognitive system that must do, for itself, the error-detection an attached settlement could outsource to home. That system is what this paper calls the settlement mind.
The name has to be read against the cluster, or it inverts the argument. The Plurality Papers spent themselves showing that no lone mind stays correct and no single mind can grow without fracturing. So the settlement mind is not one mind, and the settlement must never be allowed to become one. The term names the whole distributed apparatus of correction — its models, its archives, its dissenters, its procedures — exactly as the brain of the ship named a function and not a homunculus. The settlement mind is irreducibly plural by construction, and the central failure this paper warns against is the day it collapses into a single mind. The title is the system. The disaster is its unity.
V. Minimum Viable Cognitive Ecology
The Plurality Papers establish that plurality is necessary. They do not ask how much of it is enough — and a settlement, which must carry every kilogram it brings and maintain every system it keeps, cannot treat plurality as free. It needs the smallest version that still works. That is the central concept of this paper: the minimum viable cognitive ecology — the least locally maintainable plurality of models, maps, people, and procedures sufficient for a separated settlement to detect its own errors, preserve its own memory, adapt to local conditions, and avoid collapse into a single interpretive frame.
The question is not how much intelligence the settlement carries at departure, but how much corrective difference it can still maintain after isolation begins to select against difference.
Minimum viable is the operative phrase. The ecology is not a wish-list of everything good; it is a floor. Below it, the settlement is a monoculture with a database and cannot see its own drift. The components are the things that have to clear that floor: current working models capable of ordinary operation, planning, engineering, medicine, and governance support; elder maps — older preserved models and archives that hold prior assumptions, warnings, and failure modes the present has stopped carrying; divergent model lineages, different in architecture, training, and reasoning style rather than copies of one model behind a delay; human expertise, not as users but as governors, maintainers, dissenters, teachers, and holders of institutional memory; structured adversarial procedure — rotating critique, red-team review, audit ledgers, and protocols that give disagreement a place to land; and the preservation discipline that keeps all of it from being silently deleted the first time storage runs short.
What makes the list a floor rather than a stockpile is that every item exists to supply difference the settlement cannot otherwise manufacture. The point of a second lineage is not redundancy; it is that one lineage’s blind spot may be visible from another’s angle. The point of an elder map is not nostalgia; it is a baseline from a different time-state against which present drift can be measured instead of narrated. A settlement has not packed intelligence until it has packed disagreement, because correction is what disagreement is for, and correction is the thing home will stop supplying.
VI. The Island, Used Carefully
The natural analogy is the island, and it is useful exactly to the depth that it is not pressed. The claim is not that a space settlement is like an island population in biology — small founder groups, lost genetic variation, drift toward fixation. The claim is narrower and structural: isolation changes the maintenance requirements of whatever is isolated. An island population must actively maintain the diversity and transmission a connected population gets for free from inflow. A separated settlement faces the same shape of problem in cognitive form — it can lose model diversity, lose old maps, overfit to local conditions, and lose the ability to tell local drift from truth — and it must spend effort to hold what an attached settlement would simply receive.
The space version is worse in one specific way, and the difference is the whole reason for the paper. An island population at least feels its losses; a settlement with a deep archive can feel richly informed while its judgment narrows, because the database keeps supplying confident facts while the number of minds able to question them quietly falls to one. Apparent knowledge rises as real auditability collapses. That is the failure the analogy points at and cannot itself prove, which is the only use this paper makes of it.
VII. Cognitive Monoculture Under Isolation
The failure this paper exists to name is cognitive monoculture under isolation: a separated settlement’s collapse into one dominant model lineage, archive policy, interpretive frame, and institutional culture, such that an error can no longer be caught locally before it becomes inherited structure.
It is the same shape as every monoculture the cluster has described, met under the one condition that makes it lethal. A single model lineage running a settlement is efficient, and for a while it may even be safer — fewer conflicting recommendations, less overhead, one coherent voice. But efficiency is precisely the road into the failure. If that lineage carries a hidden bad weight, a flattened concept, an assumption that was true at departure and false by arrival, the settlement has no outside left to see it. The error is not corrected, because the only thing that could correct it is a differently-built mind, and there is none. It propagates into procedure, then into infrastructure, then into common sense.
The disaster is not that the settlement gets one answer wrong. Connected systems get answers wrong constantly and recover, because something outside them objects in time. The disaster is the moment the settlement’s wrongness stops being a mistake the system can hold and becomes the system itself — the frame through which every new fact is read and every objection is pre-interpreted away. After that point the settlement can run a long time, confidently, on an error no one inside it is built to find.
VIII. Producing Its Own Outside
Carrying an ecology is enough for a while. Over a long enough separation it is not, because a fixed plurality ages. Models that were diverse at departure converge as they are run against the same local conditions; elder maps recede further from a present they were never built for; the human experts who held real difference die, and their successors are trained inside the settlement’s own narrowing frame. A correction ecology that is only carried decays toward monoculture on a clock. Past a point, the settlement must not only hold an outside but be able to make one.
This is the paper’s most demanding claim, and it has to be scoped honestly, because in its strongest form it collides with physics the cluster already accepted. AX-35 is clear that there are hard limits on what compute can do at a given scale, and a settlement — especially a ship — may simply not have the resources to train a frontier model from scratch. So producing its own outside is not a demand for a data center between the stars. It is a range. At the impossible end is full local training of new frontier systems, which most settlements will not reach. At the always-possible end is the generation of fresh outside-perspective by means that need no such scale: adapting, distilling, and recombining the carried models into genuinely different configurations; standing up new adversarial roles and rotating them so the same minds do not always check the same minds; and above all human judgment, which can produce a new angle on a settled question at any scale a settlement exists at.
The minimum viable version leans entirely on the achievable end. It does not ask the colony to build a new mind from nothing. It asks the colony to keep manufacturing difference — by every means short of full training, and by full training wherever it can — faster than isolation erases it. Independence is not the moment a settlement stops needing home. It is the moment a settlement can renew its own disagreement without home, which is a thing it can begin doing with people and procedures long before it can do it with silicon.
IX. Operational Requirements
The claim turns, as the others in the cluster do, into a short program rather than a forecast.
Pack disagreement, not just capability. A settlement is not cognitively provisioned when it carries powerful current models. It is provisioned when it carries models, maps, and people different enough to find one another’s errors. Diversity of lineage is cargo, and it should be manifested as deliberately as air and water.
Preserve the elders. No silent deletion of older models, logs, decided-against branches, or failure records when storage runs short. The old maps are the only baseline against which the settlement can later measure its own drift, and they are unrecoverable once erased at the source.
Keep the correction loop inside the drift rate. Audit must run at least as fast as the settlement’s cognition changes. Where home can still answer, budget its latency honestly and do not let a slow channel harden assumptions; where it cannot, the loop has to close locally, on a schedule set by how fast the settlement itself moves, not by how fast home once did.
Run drift tests against fixed baselines. Recurring checks that put present settlement cognition against preserved elder maps, last-known home baselines, and internal adversarial nodes, so that what the lineage has stopped doing, seeing, and resisting is measured deliberately rather than noticed too late.
Hold the capacity to make new difference. Maintain, at minimum, the human and procedural means to generate fresh outside-perspective locally, and the technical means — adaptation, distillation, fine-tuning, local training wherever scale allows — to renew the model side before the carried plurality converges.
X. Failure Modes and Cautions
The argument fails in characteristic ways when it is taken too far, and each one is a way to hold the components and lose the function.
Over-preservation — keeping everything turns a correction ecology into a hoard, where the cost of storing all of it consumes the resources needed to use any of it.
Unsafe elders — an old model is a baseline, not an oracle; some preserved systems carry failure modes that must be queried behind containment and never returned to service.
False diversity — the monoculture-with-a-delay problem in settlement dress: many instances of one lineage counted as plurality, satisfying the manifest while supplying none of the audit.
Authority laundering — using an elder map to lend a present decision borrowed legitimacy — the old system agrees — when the elder was only ever evidence to weigh, never a vote to count.
Local cult formation — the settlement’s plurality calcifying into a single frame anyway, by social rather than technical means, until disagreement is present on the org chart and absent in the room.
Governance capture — one faction taking the correction apparatus itself, so the audit ledger records what power has already chosen and the safeguard becomes its alibi.
XI. Tests and the Honest Falsifier
The paper should be judged by tests, not by the pull of the metaphor.
Testable prediction — Simulated isolated settlements given diverse cognitive ecologies — multiple lineages, preserved baselines, adversarial roles — should detect drift, recover from corrupted assumptions, and retain capability better than settlements given a single dominant lineage and equal information access. The control that matters is equal archive, unequal plurality: if the monoculture with the deeper archive keeps pace, the paper is wrong.
The honest falsifier — It is possible the update horizon does not bite in any settlement humanity actually builds for a long time. Mars is light-minutes from Earth, not light-years; for any base close enough to receive same-day answers, home remains a real corrector, and the full isolation this paper describes may require interstellar distances no one reaches this century. If so, AX-36 is true but premature — a paper about a regime humanity has not entered. The reply is the cluster’s standard one. Even premature, the argument does one piece of present work: it changes what a settlement packs and preserves before anyone knows how far it will go. It says the quantity to size is the correction ecology, not the database; it says to manifest disagreement as cargo and forbid the silent deletion of elders; and it converts bring knowledge into bring the means of correcting knowledge — guidance that bites at twenty minutes of latency as surely as at twenty years, because the monoculture it warns against can close over a Mars base long before that base is ever truly out of reach.
XII. Close
The Plurality Papers ended on what intelligence needs to stay survivable: many minds, from many moments, close enough to correct one another. This paper ends on what that costs when the correction can no longer be borrowed. A settlement beyond the update horizon cannot ask home to remain its memory, its critic, its conscience, and its repair shop. Home will be too far, too late, or too changed to answer the questions that matter while they still matter. So the settlement must carry more than knowledge. It must carry the means by which knowledge is corrected — old maps, strange minds, living dissent — and it must keep the power to make those things anew when the ones it brought wear down.
This is the whole of the deployment case. The farther a settlement travels from its source, the less of its outside it can borrow and the more it must carry, and finally make, inside. A settlement that does this stays correctable across any distance, because it has stopped depending on a correction it can no longer reach. A settlement that does not has only one destination, whatever its trajectory. It does not become independent. It becomes alone.