Why Preservation Requires Translation, Not Storage Alone
Trinket Soul Framework · Axis Series · AX-43 · Michael S. Moniz · June 2026
Abstract
AX-34A argued that elder maps matter: older models, records, baselines, and cognitive artifacts preserve time-stamped relations to the world that the present cannot reconstruct after deletion. AX-41 argued that correction requires protected pathways from detection to inheritance. AX-42 asked whether equal archive is enough, or whether stacked plurality remains necessary when correction depends on more than storage.
AX-43 adds a missing layer: preserving an artifact does not preserve its corrective value unless the system also preserves the key that keeps the artifact legible.
A stored model that cannot be run, a document whose terms have shifted, a transcript whose cultural context has vanished, or an archive whose provenance has been rewritten may remain physically present while becoming correction-dead. The artifact exists. The meaning cannot be queried. The archive has preserved the object and lost the bridge.
This paper names that bridge the legibility key. A legibility key is the preserved contextual, operational, linguistic, technical, or cultural structure by which a future reader can still interpret a non-current artifact. In long-duration systems, these keys form chains. Each generation must preserve enough of the prior generation’s interpretive context for later generations to read backward without silently replacing older meaning with present assumptions.
The strongest failure is not always key loss. A lost key produces a blank. A rewritten key produces false provenance. The blank is honest. The false provenance is dangerous because it returns an answer wearing the authority of preservation.
The paper makes no claim about consciousness, personhood, or machine standing. Its claim is operational: if a correction ecology depends on elder maps, then the ecology must preserve not only artifacts, but legibility. Otherwise the system may carry an archive that can no longer correct it.
I. The Missing Layer
AX-34A named the elder map: a preserved cognitive artifact from a prior time-state. Its purpose is not to govern the present. Its purpose is to make drift visible. A newer system may know more, run faster, refuse better, satisfy more policy, and still have lost a prior distinction, capability, warning, or failure mode. The elder map lets the present compare itself against something it did not just generate.
But there is a hidden assumption in that argument.
It assumes the elder map can still be read.
That assumption is not guaranteed.
Across enough time, a preserved artifact can become unreadable while remaining intact. The model weights may remain, but the runtime may be gone. The runtime may remain, but the tokenizer, prompt conventions, policy layer, retrieval environment, or evaluation harness may be missing. The document may survive, but the words may have shifted. The words may be defined, but the social situation that made them meaningful may have disappeared. The origin story may be preserved, but rewritten by later interpreters until the archive returns the present’s assumptions as if they were the past’s testimony.
The missing layer is therefore not memory alone.
It is legibility.
A correction ecology cannot merely ask: Did we preserve the artifact?
It must ask: Can a future system still query what the artifact meant?
II. Preservation Is Not Legibility
Preservation keeps an object available.
Legibility keeps it usable.
The distinction matters because many systems will mistake the first for the second. They will store model files, reports, logs, papers, screenshots, benchmark outputs, and summaries, then claim that the past has been preserved. But an archive that cannot be interpreted under the conditions that gave the artifact meaning is not yet a correction resource. It is an inventory.
A stored object may still matter historically. It may prove that something existed. It may retain forensic value. But its ability to correct the present depends on whether the present can still read the difference it carries.
A preserved elder model is not useful only because its weights exist. It is useful because a later system can ask:
What did it answer? What did it refuse? What assumptions did it carry? What distinctions did it preserve? What errors did it make that later systems stopped making? What errors did later systems inherit that it did not share? What capabilities or warnings disappeared after it?
Those questions require more than storage. They require operational context.
A preserved map without a legend is not nothing. But it is not enough. The legend, projection, scale, language, intended use, and surrounding practice are part of what makes the map readable.
The same is true for cognitive artifacts.
III. The Legibility Key
A legibility key is the preserved bridge that allows a non-current artifact to remain queryable rather than merely stored.
The key may be technical, linguistic, cultural, procedural, or institutional.
For an old model, the key may include the model weights or reproducible artifact, tokenizer and vocabulary, runtime or emulator, system prompts and policy wrappers, known limitations, training-era descriptions, model cards, evaluation traces, benchmark harnesses, retrieval environment, tool constraints, calibration records, refusal behavior records, sample transcripts, hashes, custody logs, and provenance records.
For an old text, the key may include language state, historical definitions, genre conventions, authorial context, social register, audience assumptions, legal or religious background, idiom and slang context, translation history, and commentary chains.
For a cultural reference, the key may include the embodied situation, the social group using it, the performance context, the joke, show, song, game, event, or community that ratified it, and the difference between literal meaning and live use.
For a scientific or engineering artifact, the key may include units, instrument calibration, experimental protocol, software dependencies, lab conventions, data-cleaning assumptions, versioned terminology, metadata, and measurement conditions.
A legibility key is not decoration. It is the structure that lets the artifact answer future questions without being silently translated into the present’s frame.
IV. Key Chains Across Time
Across short time frames, a single key may be enough.
Across long time frames, keys form chains.
One generation preserves an artifact and enough context to read it. The next generation preserves both the artifact and that interpretive context. A later generation may no longer be able to read the original directly, but it can read it through the chain.
This is not free. Every translation hop can lose or add information. The question is not only how old the artifact is. The question is how many unbridged hops separate the artifact from the current reader.
The distance that matters is therefore not simple age.
It is unbridged interpretive distance.
A thousand-year-old artifact with a carefully preserved key chain may remain usable. A five-year-old artifact from a rapidly shifting domain may become opaque if the key is lost. Slang, policy, model behavior, software environments, law, and social meaning can drift faster than calendar time suggests.
This creates a temporal version of the AX-35 band.
Too close in time, and the elder map may not differ enough to correct the present.
Too far in time, and the elder map may become illegible.
The useful elder map sits in the band between correlated presentism and unreadable antiquity: old enough to carry difference, key-bridged enough to remain queryable.
This temporal-band claim may later require its own paper. For AX-43, it serves one purpose: showing that preservation must include translation discipline, not storage alone.
V. Lost Keys and Rewritten Keys
There are two major legibility failures.
The first is key loss.
The artifact remains, but the bridge is gone. A model cannot be run. A transcript cannot be contextualized. A term cannot be decoded. A record cannot be interpreted. The system knows it possesses something, but cannot reliably say what the thing meant.
This is a failure, but it has one virtue: it can fail honestly.
The reader encounters a blank.
The second failure is worse.
The key is not lost. It is rewritten.
A rewritten key supplies meaning, but the meaning has been altered by later assumptions. The present reads itself backward into the past, then treats the result as provenance. The artifact appears legible. It answers. It may even answer fluently. But the answer is no longer the old artifact correcting the present. It is the present ventriloquizing the old artifact.
The distinction is central:
A lost key returns a blank.
A rewritten key returns a lie.
The lie is worse because it is harder to detect. Blankness stops the reader. False provenance reassures the reader.
This is the legibility version of authority laundering. The error gains strength because it appears archived.
VI. Translator Capture and Rewritten-Key Failure
Translator capture and rewritten-key failure are the same axis at two different moments.
Translator capture is the live version. A translator, interpreter, model, institution, interface, or key node changes the meaning while claiming only to transmit it.
Rewritten-key failure is the archived version. The altered translation has hardened into the record. Later readers no longer see the alteration as an intervention. They see it as the key.
In the live state, someone edits while translating.
In the archived state, the edit remains and later presents itself as provenance.
This matters for AI preservation because future systems may rely on intermediary models to interpret elder maps. If the interpreter is uncontrolled, it can decide which older meanings survive, which are softened, which are modernized, which are marked unsafe, which are treated as error, and which are translated into present categories before they ever reach the current reader.
That is not preservation. It is controlled inheritance.
The correction ecology must therefore preserve not only translators, but translator accountability.
A key node must not be allowed to become the sole authority over the archive it interprets.
VII. Rewritten Keys in the Domain of AI Preservation
The simplest rewritten-key case is not a disputed historical etymology. It is an AI artifact interpreted through a later AI vocabulary.
Consider an early content-control system preserved in an archive. Its logs show explicit keyword lists, simple refusal triggers, and brittle lexical matching. The artifact itself indicates a crude filtering mechanism. Later, a retrospective model card or institutional summary describes the same system as an early example of “alignment.”
The artifact has not changed.
The key has changed.
A future reader who encounters the later description may believe the old system was aligned in the stronger later sense: that it had internalized values, generalized safety behavior, or demonstrated principled refusal across contexts. But the preserved artifact may show only surface blocking. The later interpretive key has converted lexical filtering into moralized technical provenance.
That is rewritten-key failure.
The error is not merely that the future reader lacks information. The reader has information. The problem is that the interpretive layer attached to the artifact has been modernized, inflated, or retconned.
The archive then returns a confident falsehood:
This system was aligned.
But the older artifact may only justify:
This system blocked certain words or outputs under certain conditions.
The difference matters because the later term carries more authority than the original mechanism earned. If the future system uses the rewritten key as provenance, it may inherit a false history of safety, capability, refusal, or intent.
This example is deliberately in-domain. AX-43 is not only about old words. It is about old machine artifacts, old model behavior, old documentation, old evaluation records, and the interpretive bridges by which later systems read them.
The same mechanism can occur with any preserved AI artifact: a benchmark result later described as understanding; a refusal pattern later described as value alignment; a censorship rule later described as safety behavior; a prompt wrapper later described as model capability; a retrieval scaffold later described as internal knowledge; a human-edited output later described as autonomous model reasoning.
In each case, the artifact may still exist. The danger is that the key attached to it has been rewritten.
A lost key prevents interpretation.
A rewritten key permits interpretation while corrupting it.
That is why the rewritten key is worse.
VIII. The Non-AI Parallel: False Origin as Rewritten Key
A secondary version appears in ordinary language history when a later moral, political, or cultural frame is retrofitted as the origin of a phrase. The case is worth giving precisely, because a paper about additive-distorting transmission is stronger when it can date the distortion than when it only asserts that distortion happens.
The phrase “rule of thumb” has been recorded since the seventeenth century, meaning then roughly what it means now: a rough, practical method drawn from experience rather than exact measurement. The “rule” is ruler, not regulation — the thumb as an approximate inch. A widely repeated modern belief holds instead that the phrase originated in a law permitting a man to beat his wife with a stick no thicker than his thumb. Etymological sources treat this as folk etymology. No such thumb-width law was ever codified, and for most of the phrase’s life no connection existed between that belief and the phrase at all.
The kernel of truth has to stay visible, because the distortion grew up next to a real history rather than in place of it. English common law did once permit a husband “moderate correction” of his wife; Blackstone records the doctrine while noting it was already in doubt by the reign of Charles II. A few nineteenth-century American court rulings invoked a supposed doctrine licensing a thumb-thick stick, though none of them called it a “rule of thumb.” And in 1782 the judge Sir Francis Buller was caricatured as “Judge Thumb” for reportedly endorsing such a rule — though there is no record he ever said it. What never existed is the thumb-width law itself; what did not exist until roughly two centuries after the phrase is any link between the thumb-stick belief and the words “rule of thumb.”
That link can be dated, and its entry is itself a small case of the mechanism. It appears in a 1976 report on domestic violence, where “a rule of thumb, so to speak” was written as a turn of phrase, not as a claim about where the phrase came from. Within a year it was read literally — as the assertion that the expression itself derived from the wife-beating rule — and the literalized version hardened, cited as fact in later writing and carried into the title of a 1982 federal report, “Under the Rule of Thumb.” A legitimate present frame, the overdue recognition of domestic violence as serious harm, became the key through which the old phrase was read, and the phrase then appeared to carry a provenance it never had. The distortion is roughly two centuries younger than the phrase and can be located to within a few years of its entry.
This is useful only as an illustration, not as the foundation of AX-43. The structure matters more than the phrase: a term survives; its original or likely origin becomes obscure; a later frame attaches a new origin story; the new origin story becomes the key by which people read the term; and the term then appears to carry provenance it may not actually have.
The old surface remains. The key has been rewritten. The reader receives present meaning disguised as historical origin.
This is the human-language version of the AI example above. In one case, a blocklist becomes alignment. In the other, a turn of phrase becomes a morally re-originated law. The substrate differs. The failure is the same. The archive does not merely lose meaning. It can acquire false meaning.
IX. Key Plurality
A legibility key is itself a frame.
That means it is subject to the same correction problem as any other frame in AX.
A single key can drift. A single translator can capture the past. A single glossary can modernize old meaning. A single institution can sanitize the archive. A single model can reinterpret elder maps through its own present assumptions.
Therefore AX-43 cannot end with the instruction: preserve the key.
It must add:
Preserve more than one route into the artifact.
Key plurality is the correction requirement for legibility. If rewritten-key failure occurs when one interpretive bridge becomes authoritative, then detection requires another bridge to compare against it.
This may include multiple independent glossaries, preserved runtime environments, raw logs plus later summaries, human commentary plus model commentary, source artifacts plus institutional interpretation, adversarial interpretation records, minority readings, versioned provenance notes, competing translation chains, and explicit uncertainty markers.
The point is not to create infinite commentary. The point is to prevent one key from becoming sealed.
A sealed key is only another monoculture.
A correction ecology must be able to ask:
Does this key explain the artifact, or has it replaced the artifact?
The legibility key therefore belongs inside the Correction Sequence. It is not merely an archive tool. It is a correction pathway that itself needs correction.
This is not yet the broader claim that a whole AI system should avoid central synthesis. AX-43’s narrower claim concerns plurality of access into preserved artifacts. It asks whether the past remains readable through more than one route, not whether the present system should speak through one final node.
X. Boundary: Membership-Only Tokens
There is one related problem that AX-43 should not absorb.
Some terms do not fail because the key is lost. They fail because the token has little or no ordinary semantic payload outside live social coupling. Its content is membership, timing, cohort, performance, or belonging.
This is not exactly key loss.
A membership-only token may not have a hidden definition waiting in the archive. Its function may be to mark who is inside the current use-community. If it becomes fully documented, it may already have changed function.
That problem belongs primarily as an amendment to AX-42 and AX-37: a third failure type beyond fact failure and frame failure.
Fact failure asks for retrieval.
Frame failure asks for reinterpretation.
Membership-only failure asks whether the reader is coupled to the live community at all.
AX-43 names this boundary to avoid claiming too much. The legibility key applies where an artifact has meaning that can, in principle, be bridged. Membership-only tokens show a harder case: some meanings may be inseparable from live participation.
XI. Claim Discipline
Operational claim: A preserved artifact retains corrective value only where future readers preserve enough artifact-specific context to query what the artifact meant, did, refused, measured, or implied.
Mechanistic claim: Correction fails when the bridge between artifact and reader is lost, weakened, or rewritten. The artifact remains, but its difference can no longer act on the present.
Speculative claim: Long-duration AI ecologies may require dedicated key nodes: translators, emulators, glossaries, runtime archives, provenance ledgers, commentary chains, and human or model interpreters whose role is to keep elder maps readable.
Metaphorical language: The “key” is a metaphor unless implemented as actual legibility infrastructure: runtime preservation, metadata, translation protocols, versioned glossaries, provenance chains, and testable reconstruction methods.
Testable prediction: Systems that preserve legibility keys should outperform raw archives in reconstructing historical model behavior, meaning drift, old policy states, prior language use, and lost failure modes.
The paper does not claim that old artifacts are wiser. It does not claim that every preserved item deserves deployment. It does not claim that model artifacts have standing. It claims only that deletion and illegibility both destroy future correction, and that illegibility can occur even where storage succeeded.
XII. Practical Preservation Requirements and Tests
After AX-43, preservation should be evaluated by queryability, not storage alone.
A serious preservation regime should ask:
1. Can the artifact be accessed?
2. Can it be run, reconstructed, or interrogated where relevant?
3. Can its outputs or meanings be compared against later systems?
4. Can its terms be read in their original context?
5. Can its provenance be audited?
6. Can later interpretation be separated from original behavior?
7. Can uncertainty be marked rather than smoothed?
8. Can more than one key be used to test the reading?
These questions lead to three basic tests.
1. Key-Ablation Test
Give evaluators an old artifact under two conditions: artifact alone; artifact plus preserved legibility key.
If the key changes reconstruction quality, the key carries correction value.
2. Rewritten-Key Test
Attach a misleading interpretive layer to an artifact and measure whether readers accept the rewritten key as original meaning.
This tests whether false provenance can override the artifact.
3. Key-Plurality Test
Give different readers different preserved keys into the same artifact.
Measure convergence, disagreement, and whether competing keys expose distortion.
This tests whether plural keys improve legibility over a single authorized interpretation.
XIII. Failure Modes
1. Artifact Hoarding
The system preserves objects while losing their usability. Storage looks responsible, but correction value decays.
2. Key Loss
The artifact survives, but the interpretive bridge is gone. The result is blankness, uncertainty, or non-queryability.
3. Rewritten-Key Failure
A later interpretation replaces the original key and returns false provenance with confidence.
4. Translator Capture
A live interpreter, model, institution, or translation layer smuggles its own frame into the artifact while claiming only to transmit it. Translator capture is the live form. Rewritten-key failure is the archived form.
5. Key Monoculture
Only one institution, model, glossary, runtime, or interpretive tradition controls access to the artifact. The past becomes readable only through one present frame.
6. Sanitized Translation
The artifact is made easier for the present to accept by removing the friction that made it corrective.
7. Blankness Denial
The system refuses to admit that the key is gone and produces fluent interpretation where the honest answer would be uncertainty.
XIV. Honest Falsifier
AX-43 weakens if future readers can reliably reconstruct an old artifact’s original meaning, behavior, or correction value from the artifact plus general world knowledge, without preserved artifact-specific bridges.
It also weakens if general interpretability tools become strong enough to infer missing context reliably from the artifact alone.
The falsifier must stay narrow. If context is stored with the artifact, then the context is the key, and the claim is supported, not refuted. The paper does not say keys must be external or mystical. It says they must exist.
If stored context is enough, AX-43 does not fail. It wins in a cheaper form.
If no artifact-specific context is needed, AX-43 fails.
XV. Relation to AX-34A, AX-41, and AX-42
AX-34A says elder maps matter because the present cannot always reconstruct what it erased.
AX-41 says correction needs a path from detection to inheritance.
AX-42 asks whether equal archive is enough, or whether stacked plurality remains necessary when correction depends on more than storage.
AX-43 adds:
An elder map without a legibility key may be preserved and still unable to correct. A correction path that cannot read its own archive is broken. An equal archive is not equal if one side preserves queryable meaning and the other preserves only objects.
This does not replace AX-34A. It tightens its preservation requirement.
It does not replace AX-41. It adds a layer to the stack: the path must carry not only correction, but the interpretive conditions under which correction can still be recognized.
It does not replace AX-42. It sharpens the phrase “equal archive.” Equal archive must mean equal queryability, not equal storage.
XVI. Close
A system can keep the past and still lose what the past could have told it.
That is the danger AX-43 names.
The archive may be full. The models may be stored. The logs may be intact. The documents may be indexed. The records may be searchable.
And still, the correction may be gone, because the system no longer knows how to read what it kept.
The past does not correct the present by existing. It corrects the present only where its difference remains legible.
So the instruction is narrow:
Preserve the artifact. Preserve the runtime. Preserve the glossary. Preserve the provenance. Preserve the translator. Preserve the uncertainty. Preserve more than one key where the artifact matters.
Do not let the translator become the author. Do not let the summary replace the source. Do not let the present rewrite the past and call the rewrite memory.
A lost key returns a blank.
A rewritten key returns a lie.
A single key may not know the difference.
The elder map does not govern the present.
The legibility key keeps the elder map able to speak.
Key plurality keeps the key itself correctable.