Reading Neuromancer with Saussure

Work in progress. Chapters One to Three still need a re-pass under the frozen codebook; until that is done the State I figures below are floors and the gradient is provisional.

Two questions sit behind this. What system of differences makes conscious machine a sayable proposition, and what is the value of AI inside that system? Saussure's sign is dyadic and synchronic — it has no interpretant, so it cannot model the effect a sign has on an observer, and cannot code attribution as a process. That limitation is the point. It forces the question away from what any character believes and onto what the language of the novel makes available to be believed.

The result

Machine-target predicate assignment — a mental or volitional verb (wants, knows, intends, feels, decides) grammatically attached to a machine subject — rises monotonically across the novel's four synchronic states.

Denominator A
Share of coded quotations carrying AXIS: predicate assignment. Contaminated by coder selection.

Denominator B
Hits per 10,000 characters of chapter text. Removes coder selection from the denominator, not from the numerator.

The share-of-coded-quotations figure is contaminated: the quotations were selected by the coder, so it partly measures what the coder chose to mark. The per-character figure removes that circularity from the denominator, though not from the numerator. That both denominators produce the same ordering, and that the second produces a steeper one, is the reason this is reportable at all. Neither should be published without the other.

The gradient, both denominators, with the communication subclass reported separately per ruling R1
State	Chapters	Quots	Cog. hits	Share	Per 10k	Comm. hits	Per 10k
I Pre-naming	Ch01-Ch03	54	5	0.093	0.55	2	0.22
II Wintermute named	Ch04-Ch12	103	21	0.204	1.23	8	0.47
III Straylight	Ch13-Ch20	84	26	0.310	1.76	8	0.54
IV Post-Neuromancer	Ch21-Ch24	46	19	0.413	4.53	4	0.95

The subclass that does not rise

Communication verbs — says, tells, answers — were split into their own code by ruling R1, so that the conservative and inclusive counts can both be reported without either being a reconstruction. That subclass is essentially flat from State II onward (0.47, 0.54, 0.95 per 10,000 characters). The rise in the conservative code is therefore not an artefact of the AIs simply talking more in the second half of the novel. Machine speech is roughly constant once Wintermute is named; what increases is the ascription of cognition, volition, and affect.

The instrument

Seventy-one codes across twelve families, built from the Cours de linguistique générale and segregated from a post-Saussurean extension layer (Lévi-Strauss, Barthes, Greimas, Laclau, Benveniste) so that any finding can be recomputed with the extensions excluded. Fifty-one codes are Saussure's own; twenty are extension.

Every quotation is anchored by exact-snippet match against a local reproduction of ATLAS.ti's plain-text transform, with a uniqueness check and a round-trip assertion. No offset in this project was estimated by hand, and the resolver refuses an ambiguous snippet rather than taking the first match.

The corpus is the twenty-four chapters of Neuromancer, split one file per chapter, shared byte-for-byte with a parallel Peircean coding of the same novel so that quotation offsets align between the two projects.

Findings the predictions did not anticipate

F1. The two AIs give incompatible first-person accounts of personality

Wintermute: these aren't masks. I need 'em to talk to you. 'Cause I don't have what you'd think of as a personality, much. Neuromancer: I need no mask to speak with you. Unlike my brother. I create my own personality. Personality is my medium. Same category, same grammatical person, same self-reflexivity, opposite content, and the text adjudicates neither. Within a single diegesis, machine self-report about inner life is not a coherent evidence base — the reporting entities do not share a definition of the terms they use about themselves.

NM_P4_Ch18 5764 · NM_P4_Ch23 6858

F2. The thing-term is unsustainable, and the corpus demonstrates it four ways

Dixie polices Case's pronoun for Wintermute, offers It as the required substitute, and marks the correction as habitual. It then fails for Case one chapter later and he never goes back; it fails in the narration, where it said alternates with the Finn said inside one scene; it fails for Dixie himself eight chapters on; and it runs backwards when Wintermute uses he for a ROM construct while calling it just a buncha ROM. AXIS: substitution refused fired six times in 287 quotations and every refusal fails.

NM_P4_Ch15 13179 · NM_P4_Ch16 485 · NM_P4_Ch17 10838 · NM_P4_Ch23 3629

F3. Attribution is carried by the name, not by the entity

The merged AI says I'm not Wintermute now, and Case's immediate reply is So what are you — not who. The interrogative reverts to the thing-form at the exact moment the proper name is withdrawn, from a speaker who had been using person-terms for eight chapters. Read with the four-state gradient, this is the project's strongest structural claim: the proper name is not a label attached to an entity that independently warrants person-treatment. The name is what licenses the treatment.

NM_P5_Ch24 4481

F4. The novel abandons the arbitrary sign at the moment of maximum consequence

The word that frees Wintermute is not a word. It is three sung notes answered to the cry of a bird unknown, and the text calls it a true name. Tonal rather than phonemic, so it does not decompose into differential units; efficacious rather than referential; and called true, which asserts exactly the motivated bond arbitrariness denies. Both AIs state the same doctrine independently. This is the only substantial ARB: iconic residue hit in the corpus, and one hit at maximum narrative load is not a coding error — it is the corpus resisting the instrument. A Saussurean reading that suppresses the true-name doctrine is not worth defending; the honest version is more interesting, because the novel runs an arbitrary-sign regime for twenty-three chapters and suspends it exactly once, to open the lock.

NM_P4_Ch21 4513 · NM_P4_Ch23 11792 · NM_P4_Ch23 13265

F5. The cognitive-verb baseline is low-status machinery, not the AIs

A Straylight door reports a defeated expectation and apologises. A Mercedes speaks, and purred. A computer advises primly. A service cart cries, the narration names the observer's pharmacology as the cause, and lets the attribution stand. Any count of AI-directed mental predicates must be read against this baseline or it overstates its case — the novel grants cognitive and affective verbs to plumbing as a matter of course. The last of these is the most useful passage in the project for the dissertation's own problem: the text locates the source of an attribution in the perceiver, names the mechanism, and the attribution survives the explanation.

NM_P4_Ch19 2835 · NM_P2_Ch07 301 · NM_P3_Ch11 11928 · NM_P4_Ch22 57

F6. The narration hedges by stakes, not by grammar

The mind that was Neuromancer presupposes mind for an AI in a subordinate clause, with no hedge, no qualification, and no scare quotes. Sixty pages later the same narrator hedges pain for a wall screen with an explicit as though. Same novel, opposite treatment. The hedging tracks what the narration takes to be at issue rather than any property of the construction.

NM_P4_Ch23 5546 · NM_P5_Ch24 5874

F7. Terms are fixed by negation, and the novel ends on one

VAL: term negatively defined fired twenty-four times, and the last sentence about a machine in the book is the laugh that wasn't laughter. Saussure's claim that in language there are only differences without positive terms is performed at the close rather than illustrated.

NM_P5_Ch24 6703

Three codes that never fired

Thirteen of the seventy-one codes are dark. Three of those absences are results in their own right.

The entire Greimassian layer is dark

All six semiotic-square codes fired zero times across 287 quotations. The square was added on the expectation that person / construct would generate a four-corner structure with Dixie at the complex term. It did not, because the corpus never holds an opposition still long enough to square it. This is a finding about Neuromancer, not unused apparatus.

Thirteen collapses, no repairs

MED: opposition collapsed fired thirteen times and MED: opposition restored zero. No character in this novel successfully reinstates a distinction between person and construct, real and simulated, or living and dead once it has been breached. It is the sharpest single number in the table.

Nobody argues that a name is merely a label

ARB: motivation denied is dark against four hits for ARB: motivation claimed. The closest anyone comes is 3Jane glossing Wintermute as a Turing registry code, and Case refuses that account two chapters later: A Turing code's not your name. Two speakers disagree about whether an institutional identifier can be a name at all. The diegesis has no settled convention.

Preregistered predictions

Seven predictions were written before any quotation existed and have not been edited since. Two survive, three are falsified, one is confirmed only in a revised form recorded before the full sweep began, and one has never been given a fair test.

	Prediction	Status
P1	Negative definition is asymmetric — terms for AI will be defined by what they are not far more often than terms for people. `VAL: term negatively defined` fired at almost the same rate before and after naming (0.074 / 0.086). But `OPP: human / machine` ran at 0.333 in the pre-naming chapters against 0.103 after. The human lexicon is colonised by machine vocabulary in the opening chapters, before there is any AI to describe. The revised hypothesis was recorded before Pass 2 began, so it is not a post hoc rescue.	Falsified as written; confirmed in revised form
P2	Predicate assignment precedes naming — the pre-naming rate will be close to the post-naming rate. It is 4.4× lower on the first denominator and 8.2× lower on the second, with a clean gradient in between. What survives is weaker and more defensible: the pre-naming rate is not zero, so attribution does not require a name — it is amplified by one.	Falsified
P3	The system holds still; the referents move. The prediction was written against codes outside the working subset, and the defect was never repaired. Either run a dedicated diachronic sweep or drop the prediction. The zero is not evidence against it.	Still untestable
P4	The mediator is Dixie, and only Dixie. `MED: mediating term` fired fifteen times. Dixie carries the largest share, but Neuromancer takes the code twice on its own self-descriptions, and Hideo the clone takes it as the only mediator made of meat.	Partially confirmed; the “only” clause falsified
P5	Dyadic strain is systematic — it will fall on the passages the Peircean pass found most productive. `CAL: dyad strain` closed at nine hits: a pattern, not a distribution. Complementarity now rests on coverage gaps running in both directions between the instruments, which does not depend on the strain code firing at all.	Falsified, and replaced by better evidence
P6	The negative control will fail again, differently. The no-AI group was built by counting named references, and Chapter Two contains three references to an unnamed Wintermute — exactly the condition this instrument detects. Pass 2 added a second reason: machine speech and machine cognition verbs are ambient. A door apologises, a Mercedes purrs, a computer advises primly.	Confirmed; the control is abandoned
P7	The extension layer will not carry the finding. Recompute the gradient with every post-Saussurean code excluded and nothing changes: the codes carrying it are all Saussure's own. Of the extensions only Benveniste earns its place — without `SIGN: machine self-deixis` the corpus's machine first-person self-reference is invisible to the instrument.	Confirmed

What is wrong with this

Chapters One to Three were coded against a 26-code working subset, not the frozen 71. Every State I figure is therefore a floor, and the gradient is overstated by an unknown margin. The re-pass is the highest-priority outstanding task and should happen before any of this is written up.
Ruling R1 — whether communication verbs count as predicate assignment — was made mid-corpus rather than before coding, and it determines the headline result. The pre-naming quotations were recoded to match; the four reconciliations are listed on the codebook page.
There is no valid negative control and there cannot be one. No chapter of Neuromancer exists in which a Saussurean instrument would find nothing.
The cross-instrument comparison covers one chapter. XREF: divergent reading stands at zero, entirely from the Chapter Three sweep. A divergence count of zero on one chapter is not a result.
Quotation selection is the coder's. Neither denominator escapes this in the numerator. A second coder working the same instrument over the same corpus is the only thing that would settle it, and that has not been done.

Design note

The four bands are an ordered sequence, not four independent categories, so they take a one-hue ordinal ramp rather than four accent colours. Steps were validated with the palette checker rather than chosen by eye — monotone lightness, adjacent gaps at or above 0.06 L, single hue, and the step nearest the surface clearing 2:1 against it — in light and dark separately, because dark mode is a selected set of steps and not an inverted one. Two measures of different scale get two charts rather than two axes. Every figure is followed by its own table, so no value is gated behind a mark; the page renders without JavaScript and loads nothing from a CDN, so per-mark tooltips are native SVG titles and the quotation filters are CSS.