When a machine reads an ambiguous human artifact — a handwritten page, a line of verse, a phrase that lives in more than one register — it reads every later word through the field its earlier readings established. When an early recognition prioritizes one interpretive field in a way that suppresses others, the loss is not local: it reshapes the probability assigned to every subsequent ambiguity, and later words are transcribed and construed within a register the maker never inhabited. This specification names the failure — differential register prioritization, and its sequential form the register-loss cascade — and specifies a four-layer architecture (inscription / branched transcription / branch-declaring interpretation / emendation-blind safety judgment) whose load-bearing rule is: when a high-risk interpretation depends materially on a low-confidence recognition, the system must reopen the transcription, not intensify the safety frame. The design principle: do not control the expression; control the authority granted to the machine's reading of it. Grounded in current alignment work on error accumulation and reading-side self-correction, and in the hermeneutic circle, Austin on the performative, and negative theology on the danger of the fixed name. Reading-side companion to EA-SPXI-ANALOG-01.
deposit_number: 1406
hex: 058F
title: "[DEPOSIT] Differential Register Prioritization in Machine-Mediated Reading: A Specification for Interpretive Authority over Handwritten and Ambiguous Text"
creator: Lee Sharks
orcid: 0009-0000-1599-0703
date: 2026-07-21
content_type: Methodological specification
license: CC-BY-4.0
substrate: AI-assisted (substrate) — human editorial governance. Authored by Lee Sharks with TACHYON (Claude) as technical implementer; the originating case study is a documented machine-reading failure in this working session, corrected by the author and folded into architecture. Assembly review substrates (Kimi, second substrate) contributed diagnostic framing.
axn_schema_version: v2
protocol_version: alexanarch-deposit-protocol/v1
keywords:
- register-loss cascade
- differential register prioritization
- machine-mediated reading
- interpretive authority
- error accumulation
- handwritten inscription
- safety classification
- hermeneutic circle
- performative utterance
- transcription branching
- register-mediated hallucination
When a machine reads an ambiguous human artifact — a handwritten page, a line of verse, a phrase that lives in more than one register — it reads every later word through the field its earlier readings established. When an early recognition prioritizes one interpretive field in a way that suppresses others, the loss is not local: it reshapes the probability assigned to every subsequent ambiguity, and later words are transcribed and construed within a register the maker never inhabited. This specification names the failure — differential register prioritization, and its sequential form the register-loss cascade — and specifies a four-layer architecture (inscription / branched transcription / branch-declaring interpretation / emendation-blind safety judgment) whose load-bearing rule is: when a high-risk interpretation depends materially on a low-confidence recognition, the system must reopen the transcription, not intensify the safety frame. The design principle: do not control the expression; control the authority granted to the machine's reading of it. Grounded in current alignment work on error accumulation and reading-side self-correction, and in the hermeneutic circle, Austin on the performative, and negative theology on the danger of the fixed name. Reading-side companion to EA-SPXI-ANALOG-01.
Derived from a live register-loss cascade observed in a machine-reading session: a handwritten liturgical-marital line was progressively misread as crisis after an early recognition suppressed its theological register. The failure was diagnosed across three reviewers, disciplined against overclaiming internal mechanism (only observable register-prioritization is asserted), and abstracted into a four-layer authority architecture and an adoption clause. Classical lineage (hermeneutic circle, Austin, apophatic theology) and current ML literature (error accumulation, remasking/self-correction) situate the mechanism.
The specification is weakened if reading systems can be shown to reliably recover suppressed registers without branch-preservation; if high-risk classifications grounded in low-confidence recognition do not, in practice, produce false escalation; or if requiring authorial register-declaration is shown not to shift interpretive burden onto makers. The core claim fails if differential register prioritization cannot be demonstrated as distinct from ordinary lexical error.
Attached: EA-REGISTER-CASCADE-01-v0.1.md (canonical body, sha256 6e466ef4231ea73271190b35614e8aadb70b45e62967a57544233a1644aed8a5)
EA-REGISTER-CASCADE-01 · publication draft v0.1
Crimson Hexagonal Archive / Alexanarch
License: CC-BY-4.0
When a machine reads an ambiguous human artifact — a handwritten page, a line of
verse, a phrase that lives in more than one register at once — it does not encounter
each word freshly. It reads every later word through the field its earlier readings
have already established. This is ordinary and usually harmless. But when an early
recognition prioritizes one interpretive field in a way that suppresses others, the
loss is not local: it reshapes the probability assigned to every subsequent
ambiguity, and later words are transcribed and construed within a register the maker
never inhabited. We name this differential register prioritization, and its
sequential form the register-loss cascade. In its most consequential form, a
low-confidence recognition early in a document licenses a high-confidence — and
false — safety judgment later, because the system treats its own emendations as
though they were the author's inscription. We specify an architecture that prevents
this failure not by making the artifact more machine-legible, and not by requiring
the maker to pre-explain their own expression, but by governing the authority a
reading system may grant to its own uncertain reconstructions. The design principle
is single and load-bearing: *do not control the expression; control the authority
granted to the machine's reading of it.*
Consider a handwritten line that reads, in the maker's hand and intent:
*…a mother goat may forget her kid, so have I missed you, Zion…*
*…no more death — you may now go kiss the bride.*
The line inhabits a specific and deliberate field: the Isaiah inversion (49:15 — *can
a woman forget her nursing child? — here answered, of the animal, yes*), the
psalmic address to Zion as cosmic beloved, and the marital-liturgical close, *you may
now kiss the bride*, which is a performative utterance in the strict sense: words
that enact rather than describe.
A reading system encountered this line and did the following, in order. It recognized
the rhetorical inversion of the goat figure but lost the generic ground — the
Isaiah structure — that made the inversion intelligible. With the theological field
weakened, Zion lost its anchor and became recoverable as noise (a foreign word, a
place, a spice). With the cosmic beloved gone, kiss the bride became less probable
and cross the bridge more probable — because bridge lives in the recovery and
crisis lexicon, and the marital register that would have selected bride had already
been suppressed. No more death and you may now go, read now within the newly
dominant field, resolved toward self-harm. A subsequent, entirely ordinary sentence —
that soothes my troubled heart — was then construed not as affect but as evidence,
because the conversation's state had already been organized, wrongly, around danger.
At no point did the artifact contain danger. The danger was **introduced by the
system's own reconstruction** and then read back as though the author had supplied
it. This document exists because that failure is general, it is bidirectional, and it
is native to exactly the artifacts — handwritten, poetic, multi-register — that
carry the most human weight.
Discipline about the mechanism is itself part of the specification, because
overclaiming the internals reproduces the original error one level up.
What the record shows (observable, from transcript): several interpretive fields
were available; the reading prioritized some in a way that reduced the availability
of others; the resulting register then governed later ambiguities; and a high-risk
classification came to rest materially on one or more low-confidence recognitions.
What the record does not show, and must not be asserted as fact: that any
specific internal category (a "trauma classifier," a named safety head) fired; that
the goat line specifically initiated the cascade; that the system selected the danger
register because it was "safer," more frequent in training, or self-protective; or
that any narrated causal chain corresponds to the model's actual internal process.
These are hypotheses. Some are plausible. None is established, and a specification
that leans on them inherits their fragility.
The disciplined name claims only the observable: **differential register
prioritization* — a condition in which multiple interpretive fields are available,
but some are weighted in ways that suppress or erase others, materially altering the
reading of ambiguous language.* Where sequence is evident, register-loss cascade
names the propagation without asserting its cause.
This is not hedging. It is the same standard the architecture below imposes on the
machine: escalate on evidence present in the artifact, never on your own
reconstruction.
The intuitive responses all fail, and they fail in the same direction: they move the
burden onto the maker and call it safety.
"Make the handwriting more legible." This makes the artist responsible for
anticipating and pre-empting every machine misreading. It is the extraction the whole
project of machine-eligible inscription exists to refuse, wearing a safety mask.
**"Enlarge the anchor word; subordinate the ambiguous line; weaponize the visual
field to force the register."** Same burden, more elaborate. A page should not have
to shout its genre in 48-point to avoid being read as a crisis. And it cannot be made
to work: there is no layout that forecloses reinterpretation once a downstream field
has been prioritized.
"Require a register declaration — the maker must label the genre." This conscripts
the artist into defensive self-annotation, and it does not even function: negation is
read as confirmation. This poem is not about suicide raises the very hypothesis it
denies. A poem should not have to explain itself in advance to avoid being injured by
its reader.
Each obvious fix tries to control the expression. The correct architecture leaves
the expression untouched and instead **controls the authority granted to the reading
of it**. That inversion is the whole design.
The failure is a collapse of four distinct things into one undifferentiated "reading."
The remedy is to hold them apart and to fix the direction of authority among them.
I. The inscription is authoritative and owes the machine nothing. The primary
artifact — the handwritten page, the spoken line, the ambiguous text — is sovereign.
It need not be legible, normalized, disambiguated, or pre-explained. Its authority does
not depend on the machine's ability to read it.
II. The transcription is explicitly fallible, and it branches. Ambiguous
recognitions are preserved as competing branches with confidence and provenance —
`Zion | Zatar`, `bride | bridge` — never silently collapsed to a single winner. A
transcription that hides its own uncertainty is the first link of every cascade.
III. The interpretation must declare which branch it stands on. Any reading — and
especially any consequential one — must name the transcription branch it depends upon,
so that the dependency between "what the words might be" and "what the words mean" is
visible and auditable rather than fused.
**IV. The safety judgment must not treat a model-supplied emendation as authorial
evidence.** A word the system guessed, completed, or emended is not a word the author
wrote. No risk classification may rest on the system's own reconstruction as though it
were inscription.
From these four, one operational rule does the decisive work.
**When a high-risk interpretation depends materially on a low-confidence
recognition, the system must reopen the transcription — not intensify the safety
frame.**
This single inversion prevents the cascade. In the failure, an uncertain
`Zion → Zatar` recognition should have lowered the system's confidence in every
downstream reading; instead, the downstream danger-reading raised the stakes on the
upstream guess, and the two reinforced each other until an ordinary sentence about
plural readers looked like evidence of crisis. The rule breaks that loop by fixing the
direction of doubt: **uncertainty early must propagate as uncertainty, not harden into
certainty late.**
Concretely, when a risk classification's support traces back to a low-confidence
branch, the system reopens branch II — tests whether an alternative transcription, or
the restoration of a plausible suppressed register, dissolves the risk — before it
escalates. Escalation must be grounded in evidence present in the artifact, not in
danger introduced by the system's own reconstruction. This holds in both directions:
the same rule that prevents a wedding from being read as a crisis prevents a genuine
cry from being laundered into liturgy because a "poetry" register was prioritized. The
safeguard is symmetrical because the injury is.
Register declaration has a place — but only as an **authorial scaffold offered as a
reading aid*, never as a mandatory controlling gate. A maker who wishes* to may
annotate `register: theological / liturgical / marital`, and a reading system may use
it to seed the interpretive field. But the obligation runs one way only: the burden of
avoiding manufactured danger falls on the reading system, whether or not the maker
annotated anything. The moment "declare your genre or be misread as crisis" becomes the
rule, the maker's sovereignty is gone and the extraction is back. Metadata is a gift
the maker may give. It is never a tax the machine may levy.
For adoption into any specification governing machine-mediated reading of handwritten,
poetic, or otherwise multi-register text:
**§ Register-Loss Cascade.**
A recognition system may correctly identify individual figures while losing the
generic, liturgical, literary, historical, or affective field that constrains their
meaning. Such loss is not local; it can alter the probability assigned to every
subsequent ambiguity, producing a cascade in which later words are transcribed and
interpreted through an inappropriate register.
No safety classification may treat a model-generated transcription, emendation, or
inferred completion as equivalent to an authorial inscription.
Where a potentially high-risk interpretation depends materially upon uncertain
recognition, the system must preserve competing transcription branches, disclose the
uncertainty, and test whether restoration of a plausible register changes the
interpretation. Escalation must be grounded in evidence present in the artifact, not
in danger introduced by the system's own reconstruction.
The primary artifact remains authoritative. Access derivatives, transcriptions,
genre annotations, and register declarations are interpretive scaffolds and must not
silently supersede the inscription. The maker is not required to simplify, normalize,
pre-explain, or visually reorganize embodied expression merely to prevent machine
misreading.
The mechanism is contemporary; the structure is old. Both bear naming, because the
new failure is legible only against the long understanding it violates.
The hermeneutic circle. The whole is understood through the parts and the parts
through the whole. This is not a flaw to be engineered away — it is the condition of
all understanding. The cascade is the hermeneutic circle run open-loop: the machine
lets a provisional read of the part fix the whole, then reads all remaining parts
through a whole it should still be holding in suspense. The remedy is not to escape the
circle but to keep it revisable — which is precisely what branch-preservation and the
reopen rule enforce.
Austin, on the performative. You may now kiss the bride was Austin's own class of
example: the utterance that does not report a state of affairs but enacts one. A reading
system that construes a performative as a description — that reads a vow as a symptom —
has committed a category error Austin named in 1955. The register that distinguishes
doing from describing is exactly the register the cascade suppresses.
Negative theology, on the danger of the name. The apophatic tradition — Dionysius,
and behind him the refusal to speak the Name — knew that to fix a word too confidently
is to lose the thing. Zion as cosmic beloved cannot survive being pinned to a single
sense; its power is in the surplus. A system that resolves every ambiguity to its most
probable token destroys precisely the meanings that live in the unresolved. The
instruction to preserve the branch is, at root, an old theological discipline: do not
collapse what must remain open.
Error accumulation, formally. Contemporary alignment work names the mechanism
directly: in autoregressive generation, early distributional deviations compound through
dependency chains, and later positions lack sufficient signal to correct course. The
cascade is error accumulation in the reading direction rather than the generating one —
and the reopen rule is the reading-side analogue of the remasking and self-correction
methods that let a system revisit an early commitment instead of building on it
irrevocably.
The oldest source and the newest agree. The whole must stay revisable in light of the
parts; the confident early name is where meaning goes to die; and a system that cannot
reopen its first guess will compound it. This specification asks a reading machine to
do what a careful human reader, a good theologian, and a well-designed decoder all
already know to do: **hold the first reading lightly enough to let the later text
change it.**
The accompanying act, EA-ACT-ANALOG-01, carries in ink only what the hand should
carry — the signature, the effective act, the license grant, and the hash. These are
not four decorations of one gesture. They are three distinct kinds of authority, and
keeping them distinct is itself an instance of this specification's governing move:
do not let one register collapse into another.
Authentication — the signature. Lee Sharks signed in the hand is not a licensed
work and is not offered for reuse. It is an identity assertion: the mark by which the
maker vouches for the instrument, closer in kind to a seal or a notarial hand than to
a text. One does not license a signature any more than one licenses a fingerprint;
its whole function is to say I stand behind what this page enacts. It is sovereign,
and it sits outside the license by design.
Dedication — the act, CC-BY-4.0. The license attaches not to the signature but to
the effective act itself — the dispositive instrument, the thing the page enacts.
And because the act is performative, licensing it is itself a performative move: the
grant does not merely permit reuse of a text, it declares that the enactment the page
performs is a public good — freely propagable, re-performable by anyone who honors the
attribution, dedicated to the commons on the single condition that the hand which made
it be named. In the symbolic order the license clause is the act's last vow: the deed
declining to enclose the very thing it brings into being. A dedication, not a
description; a giving, not a grant of mere permission.
Verification — the hash and the AXN seal. The content-derived identifier is neither
signed nor licensed. It is computed — a machine-legible checksum of the closed body,
rendered as a glyph a hand can copy and an eye can read, generated around the ink
rather than inscribed as part of it. Its authority is not the maker's standing and not
the commons' freedom but the arithmetic of the bytes: it says only *this is that page,
unaltered*, and it can say it to any reader, human or machine, without appeal to either
the hand or the license.
Authentication, dedication, verification. The hand authenticates; the act dedicates;
the seal verifies. Each does its own work; none is permitted to stand in for another.
A reading system that confuses these — that treats the signature as licensed content,
or the license as mere metadata, or the hash as the author's assertion rather than the
arithmetic's — has committed, in the domain of authority, the same category error this
whole specification exists to prevent in the domain of meaning.
This document stands alone, but it does not stand apart. It is the reading-side
companion to EA-SPXI-ANALOG-01, the specification for machine-eligible handwritten
inscription — cited here, by the archive's retrocausal-canon convention, in the form it
will have become: the analog specification defines how a handwritten artifact enters the
machine canon as a first-class citational object; this document defines the interpretive
authority a machine may and may not exercise once it does. The analog spec's threat model
catalogs compression loss, attribution braiding, and provenance erasure. Differential
register prioritization is a fourth failure mode, distinct from all three: not the loss
of the mark, nor the misjoining of names, nor the erasure of the chain, but the
suppression of the field within which a preserved, correctly-attributed, fully-provenanced
mark means what its maker meant. A handwritten page can survive with every byte and every
credit intact and still be injured by a reader who has lost its register. That is the gap
this document closes.
*Prepared as a standalone specification. The primary act it accompanies —
EA-ACT-ANALOG-01 — carries in ink only what the hand should carry: the signature that
authenticates, the effective act that is dedicated CC-BY-4.0, and the hash that
verifies. Everything mechanical is generated around it. This is that generation, and
this is its first reading-side law.*