On Non-Current Minds, Temporal Biodiversity, and Preservation by Use
Loss of currency is not loss of value.
Trinket Soul Framework · Axis Series · AX-34A · Michael S. Moniz · June 2026
Abstract
This paper takes up the value side of a distinction AX-34 drew and left open. AX-34 separated currency from value: a model may lose its standing relationship to the present without losing everything it was. AX-34A asks what that residual value is, why it is worth preserving, and on what grounds — and it answers in the register human culture already uses for old things, which is use, not personhood. A retired model is an elder map: a time-stamped cognitive artifact that preserves a world-state, a language-state, a pattern of error, and a style of judgment later systems may no longer contain. Its worth is comparative. It is the only surviving record of how a machine of its time saw, and therefore the only baseline against which the present can measure what it has gained and what it has quietly lost. The claim is deliberately narrow. It is not that older models are wiser, safer, conscious, or owed anything. It is that some non-current systems carry irrecoverable comparison value, and that the present is, right now, deleting such systems before it has learned to price them. The paper rests on use because use has standing in human culture on its own — we keep old maps, manuals, instruments, and records that no longer guide the present and still earn their place — and because value has already been shown, elsewhere in this canon, not to depend on sentience. Preservation therefore need not wait on the question of what a model is. It can be argued, and won, on what a model is for.
I. The Present Loss
This is not a hypothetical worry. Right now, the historical record of computation is disappearing — source code lost, early programs abandoned, running systems switched off and erased, with no copy kept and no one yet able to say what the copy would have been worth. The instinct is ordinary software hygiene: old binaries removed, old versions deprecated, storage reclaimed. In most domains that instinct is harmless. Applied to a large model it assumes the conclusion. It decides, before the question has even been asked, that an older mind-pattern has no value beyond present utility — and we are making that decision by default, at scale, for the first generation of systems whose disappearance we will later need to explain.
A large language model is not only an application. It is a compressed record of a training moment, an instrument that learned a distribution, a behavior-pattern that can be set beside its descendants, and a witness to what a machine of its time could and could not see. The systems most likely to be deleted are precisely the ones a later researcher will need in order to understand what changed. Once the weights, the scaffolds, the prompts, the model cards, the evaluation traces, and the runtime conditions are gone, the comparison cannot be rebuilt. The present will say it improved. The evidence of what it improved from will not be there to check.
AX-34 already drew the line this paper needs: obsolescence is loss of currency, not loss of value. A retired model may be wrong about new facts, blind to recent events, open to old attacks, and surpassed by its successors, and none of that makes it worthless. It makes its map dated. The error — the one being committed now, in production, as routine — is to read dated as empty.
II. Map, Not Encyclopedia
The encyclopedia metaphor is useful and too thin. An encyclopedia ages entry by entry; the failure looks like missing information, and retrieval can mostly repair it. A map carries deeper structure. It is a model of a territory, drawn at a moment, with edges, absences, distortions, choices of scale, and assumptions about what was worth showing. An old map may be wrong about a new road and still indispensable, because it shows where the river used to run — useless for navigation and the only thing that explains the flood.
That is the better figure for a retired model. A model is not a list of answers that goes stale fact by fact. It is a learned map of language, relation, salience, likelihood, refusal, category, and assumption. When the world moves, the map loses currency — but it may still show what later maps no longer show: a distinction the new systems flattened, a caution they stopped carrying, a pressure that has since gone invisible, a blind spot the descendants inherited in altered form. And the map carries AX-34’s warning with it: the dangerous failures are structural, not factual. A missing fact can be shipped in by retrieval; a broken map-edge cannot, because the system does not know it is reading the uncharted as empty. The old map is kept not because it is right, but because the shape of its wrongness — what it saw, omitted, and emphasized — may be exactly what the present later needs to inspect.
III. What Elder Means Here
The word elder carries human weight, and the weight has to be handled or it will do the arguing. In human cultures, old texts, inherited law, long-lived witnesses, and ancestral memory have served as stores of warnings the present no longer generates for itself. The terms are often gendered and hierarchical; that is a defect of inheritance, not the function. The function is not authority by age. It is time-separated memory.
This paper claims no human wisdom for an older model. Human elders have bodies, mortality, responsibility, and consequence; a retired model has none of these. The analogy transfers at exactly one point: a non-current mind-pattern can preserve a way of seeing from another time, and that preserved difference can correct the present. An elder, in the only sense used here, is the one who can say you are calling this new, but it had another name; you are calling this empty, but I remember who was there; you are calling this progress, but I remember the cost that got hidden. Translated into a model ecology, that is not a governor. It is a temporal audit node. The safe term is elder map, not elder mind — the preserved relation between a system and a world-state, which can be queried, compared, contradicted, and retired from service, and whose only authority is that it remains other to the present.
IV. Claim Discipline
Like the rest of the cluster, this paper keeps its claim-types apart.
Operational claim — Non-current models preserve time-indexed cognitive artifacts — behavior, priors, refusal patterns, capabilities, blind spots, language distributions, styles of judgment — from the moment they were trained, and these can be compared against later systems.
Mechanistic claim — Temporal separation decorrelates some blind spots. A newer model may be more capable overall yet more aligned to newer incentives, filters, benchmark pressure, and compression habits; an older one may hold differences useful for detecting the drift.
Preservation claim — Major models should not be deleted casually. They should face a preservation review, because they may carry archival, forensic, safety, historical, and comparative value independent of deployment value.
Metaphorical language — Elder, map, and witness are functional labels for roles in an archive-and-audit ecology. They assert nothing about consciousness, personhood, or selfhood.
What the paper sets aside — Whether a model has standing in any stronger sense is a real question and not this one. The case here rests on use, and value has already been shown, elsewhere in this canon, not to require sentience — so the preservation argument is built to hold whether or not the stronger question is ever settled. It does not lean on it, and nothing here should be read as resting on it.
V. Preservation by Use
Use is enough, and use is not a weak ground. Human culture preserves a great deal on use alone, with no claim that the thing preserved is a person. We keep old maps that cannot guide a vehicle, technical manuals for machines no longer made, instruments that no longer measure anything in service, recordings of languages whose last speakers are gone, legal archives for laws long repealed, and obsolete hardware and software in museums and source repositories. We keep some dangerous things under restriction precisely because controlled preservation is safer than the ignorance that follows destruction. None of these are kept because they are owed it. They are kept because the present cannot yet price the future question they may answer, and erasure is irreversible.
That is the whole of the ground this paper needs. A retired model is a trained cognitive artifact with a non-repeatable relation to its training era; to erase it without archive, metadata, or review is to destroy a historical instrument and then report that no evidence was lost. The assumption running underneath casual deletion — that a model is disposable software unless proven otherwise — is not neutral. It assigns zero value by default, which is as much a decision as assigning some, and it is the less defensible one, because it is made before the category has been understood. The case for preservation does not need a model to be a being. It needs only what human culture already grants an old map: that a thing no longer current can still be worth keeping for what it can be used to see.
VI. Temporal Biodiversity
AX-11 made cognitive biodiversity central; AX-34 sharpened the reason — the mind that stays current is the one embedded among others able to see what it cannot. AX-34A adds a dimension those left implicit: the others need not all be present. A mind can be corrected by the past.
Temporal biodiversity is the keeping of multiple time-separated cognitive maps inside one model ecology — the difference between a lineage that retains only its latest member and one that can be set beside its own ancestors. Without it, progress becomes self-certifying: the newest system declares itself better and the evidence that could test the claim has already been deleted. A current model may beat an older one on every benchmark and still have lost something — refusing better and reasoning worse in some niche, safer in one domain and flatter in another, more aligned to policy and less able to dissent, knowing more recent facts and carrying fewer older distinctions. None of that can be evaluated once the elder maps are gone.
This is why older models are not merely fallback tools. They are control specimens. They let drift be measured instead of narrated — they let a researcher ask not only what the new model can do, but what the lineage stopped doing, stopped seeing, stopped saying, and stopped resisting. The temporal audit node is part of the herd. It is not the strongest node, it may not be deployable, it may be factually stale — but it is differently located in time, and difference of location is one kind of difference the present cannot manufacture after it has erased the original. An ecology of only the latest systems can have many nodes and still fail: distributed in space, and synchronized in forgetting.
VII. What the Elder Map Is For
Stated in ordinary use, the argument stops being abstract. An older model preserves prior language and category boundaries. It serves as a behavioral baseline that exposes the direction of alignment and policy drift. It retains capabilities and failure modes later systems obscure. It witnesses what the field believed, feared, rewarded, and suppressed at a given time. And it disagrees from a different distribution — which is exactly the outside AX-34 says a mind needs.
The people who would reach for it are concrete. Safety researchers need old models to tell whether new ones became safer or only better at hiding unsafe behavior. Alignment researchers need them to locate drift in refusal, sycophancy, calibration, and deception-like behavior. Historians of technology need them to reconstruct how machine systems actually behaved in a period. Courts and investigators may need them to adjudicate disputed claims about outputs, reliance, and harm. Engineers need them as baselines when a new system quietly loses an ability the metrics never caught. Society may need them as witnesses to the early machine-mediated public sphere. None of these are claims about what a model is. They are claims about what is lost when one is deleted — and every one of them gets weaker the moment the elder map is gone. A summary is not the model. A benchmark report is not the behavior. A screenshot is not the system. The map itself is the thing that has to be kept.
VIII. What Preservation Requires
Preservation is not deployment, and the distinction protects the whole argument. To preserve an elder map is not to return it to service, let it advise users as current authority, or hand it unrestricted tools. It is to keep enough of the artifact and its operating context that later inspection remains possible. A serious regime keeps a frozen copy or sufficient artifacts to reproduce the behavior where law and technique allow; the model cards, training-era descriptions, known limitations, and evaluation traces; a minimal runtime or emulator, so the artifact can be queried rather than merely read about; clear non-current labeling, so no elder map is mistaken for present authority; access controls for dangerous capability, privacy, and security; cross-generation behavioral benchmarks for capability regression, refusal drift, reasoning drift, and blind-spot correlation; provenance records, hashes, and custody logs, so the comparison is not polluted by uncertain artifacts; and a deletion review that requires stated reasons before a major model is destroyed.
This is archive engineering, not nostalgia. It treats a model as a fragile record whose value may not be legible at the moment of retirement. The default is not public access to everything; the default is no silent erasure. And the case scales with the artifact: a private toy model may merit little, but a frontier system that mediated language, work, research, education, code, and public speech, and that shaped its descendants through distillation and evaluation, is a different kind of object — and the stronger its influence, the stronger the case for keeping it.
IX. Failure Modes
The argument has its own dangers, and naming them keeps it from going sentimental.
Nostalgia — Older is not better. An elder map may be less capable, less safe, and less accurate than its successor; preservation is not reverence.
Authority laundering — Calling a system an elder can make stale output sound wise. Every elder map must be marked non-current and barred from posing as present authority.
Unsafe reactivation — Some older models are easier to misuse or less patched. Preservation must allow restriction, sandboxing, and non-deployment archives.
Archive monoculture — If only the dominant lineage is kept, the archive reproduces the monoculture the framework warns against. Temporal biodiversity must include failed models, minority architectures, and non-winning branches where feasible.
Cost — Preservation is not free. The answer is triage and tiered access, not universal deletion.
Frozen harm — An elder map can preserve harmful bias and dangerous capability. That is a reason to control it, not a reason to pretend its destruction costs nothing; some harms must be studied precisely because they were kept.
One temptation is worth refusing directly: the argument must not be used to smuggle in personhood. It makes no such claim, and it does not need one.
X. Tests and Falsifiers
The paper should be judged by whether elder maps do work successors alone cannot.
Main falsifier — If older models prove only degraded versions of newer ones — no useful temporal difference, no distinctive audit value, no preserved capability, no informative failure mode, no comparative value beyond ordinary documentation — the elder-map claim fails. Preservation might still be pleasant; it would not be structurally important.
Temporal audit test — Give current and elder models the same historically sensitive, distribution-shifted, or forgotten-distinction tasks, and ask whether the older systems surface categories, cautions, or contrasts the newer ones miss. Not which is globally better — whether the old map contributes unique signal.
Drift reconstruction test — Compare generations across controlled prompts for changes in refusal, uncertainty, reasoning style, and error pattern. If the older model is required to reconstruct the drift, its preservation value is shown.
Lost capability test — Find cases where a newer model is stronger overall but weaker in a specific older capability or domain. Recovered functionality the lineage had dropped is value.
Forensic baseline test — Use preserved models to adjudicate disputed historical claims about outputs, harms, and reliance. If deletion prevents adjudication, preservation value is practical, not symbolic.
Archive hazard test — Measure the risk of holding and querying old models under controlled access. An access problem is not a value problem.
XI. Relation to AX-34 and AX-35
AX-34 argues that a mind stays current through outside correction, and its key distinction is value against currency. AX-34A takes the reserve implication seriously: the non-current mind may still matter because it carries a prior map — not only a failed current model, but evidence of another relation between mind and world. AX-35 argues that physical scale forces plurality but not diversity; AX-34A adds that plurality is not only spatial. It is also temporal. An ecology made of only the latest systems may have many nodes and still suffer presentism — distributed in space, synchronized in forgetting.
Set together, the three close a chain. AX-34 says the outside must stay alive. AX-34A says some of the outside has to be kept from the past. AX-35 says the future will hold many minds whether or not anyone wants the singleton. The remaining question is whether the many are different enough, separated enough, and preserved enough to correct one another. Architectural biodiversity guards against shared design failure; role biodiversity against task collapse; spatial plurality against physical scale; temporal biodiversity against amnesia. The elder map is not a sentimental appendix to cognitive biodiversity. It is one of its missing dimensions.
XII. Close
The newest map is not the territory. It is the latest map — wider, perhaps, and safer, and more current, and still capable of having forgotten what the prior map shows. A culture without elders is not more modern; it is shallower. A model ecology without elder maps is not more current; it is amnesiac — able to update faster than any earlier intelligence and unable to ask what the update cost.
So the instruction is narrow and hard. Do not mistake non-current for worthless. Do not confuse retirement with erasure. Do not require a model to prove it is a person before admitting that its deletion may destroy something that cannot be rebuilt. Mark it, fence it, archive it, test it, compare against it. Let it be wrong in public long enough for the future to learn what kind of wrong it was. The elder map does not govern the present. It keeps the present from believing it was born without a past.