The Split Ratio¶

A term, and the practice it names — fixed within the SDC canon.

Status: Canonical lexical document. Version 1.1, May 2026. PDF: The Split Ratio (PDF) — same content, with a public preface for offline reading and citation. Display your split ratio: The configurable badge lives on the Declare page.

This document fixes a term and the practice it names. The term came together through several refinements, set aside for the reasons recorded here:

nisus ratio — a Latin neologism in a vocabulary that otherwise stays in plain English; it read as out-of-family in SDC's syntactical context
tuning ratio — confused the catchy word with the term; tuning is what the practitioner does, not what the relation is called
discipline ratio — abstract, lossy; doesn't name the architectural move
split-domain ratio — accurate but heavy; reads as an attribute of the discipline rather than as a separable concept
the SDC ratio — too generic; could be misread as the ratio that defines SDC, which is false (SDC is the architecture; the split ratio is one declarative property of an artefact built under it)

What survives is recorded here.

The split ratio is an SDC concept. It belongs to the Split-Domain Cognition canon and sits alongside SDC's other foundational marks. Any instantiation of SDC inherits the ratio. The concept is not the property of any single project that builds on SDC; it travels with the discipline wherever the discipline travels.

The phrase¶

The split ratio.

Used with the definite article when referring to the relation as such. Used indefinitely when referring to a particular instance. The phrase is a noun phrase. It names a property of a single artefact: the degree to which the work holds two unlike cognitive domains apart.

Etymology¶

Split — from Middle Dutch splitten, to cleave or part, cognate with Old Frisian splita and Middle Low German splīten. The verb's primary sense is the clean separation of one thing into two — a wedge driven, a fibre divided, a path forked. In English, split arrives unburdened by metaphysical baggage: it does not pretend to elegance, and the architectural move it names is plain. Where the older Latin candidate (nisus) brought a press, the Germanic word brings a cleaver. The SDC discipline is closer to the second: not a sustained leaning, but a clean line drawn where two unlike domains would otherwise blur.

Semantics¶

The split ratio names the degree to which a single artefact maintains SDC's split-domain discipline — the structural separation between language work and judgement work, or between any two unlike cognitive domains that the artefact's context puts at risk of collapse.

The ratio is expressed as a proportion N:M on a scale of ten, where N is the share of the work that holds the split and M is the share that conflates. The two poles are split and conflated. A high split ratio means the architecture is visible in the artefact — a reader can point to where description ends and verdict begins. A low split ratio means the moves blur — description and judgement are folded into the same sentence, and the reader has no place to disagree.

The split ratio is self-rated. Unlike an alignment declaration (which makes a structural claim that can be tested against the four-test protocol), the split ratio is a self-reflective frame: this is how cleanly I am holding the discipline in this artefact, as honestly as I can read myself. It is a journal entry, not an audit verdict. Other readers may disagree; that disagreement is itself useful information. The badge does not police; it makes the practitioner's own self-reading visible.

Everything in this document, where not otherwise marked, describes the self-rated register. The second register — in which an instrument observes a conversation in real time and surfaces a provisional reading — is named sensed split and is treated separately below (see Two registers). The two registers are distinct, independent, and additive; they share vocabulary but not status.

The endpoints of the scale are not configurable. The values 1 through 9 are the meaningful positions; 0 and 10 are excluded by design. A claim of 10:0 — perfect split — invites self-flattery and is structurally suspect (no real artefact achieves clean separation everywhere). A claim of 0:10 — no split at all — would mean the artefact carries no SDC discipline whatsoever, in which case the badge has no business being attached to it. The configurable range is 1:9 through 9:1.

Relation to existing terms in the SDC lexicon¶

SDC names the architecture. Variants name the domain-specific applications. The forms of follow (support, reference, alignment, derivation) name the modes by which an external project declares its relation to the canon. The split ratio names a different thing again: the degree to which a single artefact carries the discipline, claimed by its maker.

A piece of work might carry both an alignment declaration and a split ratio. The alignment declaration claims, structurally, that the work performs the separation. The split ratio adds the practitioner's honest reading of how cleanly. The two are independent and additive; declaring one does not commit you to declaring the other.

Linguistics of use¶

The phrase functions as a noun phrase. Use the definite article when referring to the relation as such; use indefinitely when referring to a particular instance. Plural: split ratios.

Canonical sentence forms¶

The split ratio of this essay is 7:3.

The artefact-level self-report. The practitioner has read their own work and declares the proportion at which the discipline holds.

This paragraph conflates more than the rest of the piece.

The local self-report. Within a single artefact, the ratio can vary by section; a 7:3 average can hide a 9:1 opening and a 4:6 middle.

Different artefacts carry different split ratios.

The comparative claim. Two works by the same author may externally look similar and be doing different practices, because one holds the discipline and the other does not.

Tuning the split is the practice.

The definitional claim. The practice is not arriving at a correct ratio; it is the ongoing work of noticing and adjusting.

Compatible collocations¶

holding the split — sustaining the discipline against the pull of conflation
the drifting split ratio — changing without the practitioner's volition, often under rhetorical pressure
reading the split — the act of self-assessment
the close split ratio — proportion near parity (the practice in danger of collapse)
the wide split ratio — proportion far from parity (the practice held)
retuning the split — restoring the discipline after drift

Incompatible collocations¶

the perfect split ratio — implies 10:0 is a real position; it is not. No artefact is wholly split.
measuring the split ratio — the ratio is self-read, not measured. Reading, sensing, declaring are the operative verbs.
the optimal split ratio — refuses optimisation register. The ratio is to be honestly named, not maximised against any objective.
the average split ratio of the field — refuses aggregation across practitioners. The ratio is a self-frame, not a benchmark.

Register¶

The phrase belongs in writing about practice, in self-report, in journals, and in conversation with others who are also cultivating the literacy. It does not belong in productivity tooling, in scoring rubrics, in employer-facing dashboards, or in any context that converts the self-frame into an external metric. The badge declares; it does not score.

Visibility — the badge¶

The split ratio's visualisation is a divided string. A horizontal line with eleven tick marks: the two anchors (split, conflated) and nine interior positions. A sepia bridge sits at the chosen position. The bridge moves toward whichever side the artefact tilts.

The badge's title word is SPLIT — the catchy word, the architectural move. The canonical term split ratio sits below the title in italic, so the term and the practice both travel with the badge. The numerical ratio N:M sits in the right corner.

The visual register is constrained even as the form is open. Operative verbs across every implementation: reading, holding, retuning, sounding. Refused verbs: measuring, scoring, grading, optimising. A bar chart, percentage display, dashboard gauge, or productivity meter would each import the optimisation register the term refuses.

The configurable badge lives on the Declare page, alongside the other declarations the site supports.

Self-rating, not policing¶

The split ratio is, in its full meaning, a frame of self-attention. To declare one's split ratio is to claim a relation to the discipline in writing, where the writing itself is the act of holding the relation visible. The badge is a public surface for that private self-reading.

The badge does not need to be true to anyone but its declarer. A reader who finds the artefact's actual split lower than the declared ratio can say so — and that disagreement is a useful artefact of the discipline at work. But the badge is not a verdict and cannot be revoked by an arbiter. It is the practitioner's record of where they stand, today, in the holding of the architecture.

Two registers: self-rated and sensed¶

The phrase the split ratio, used as a canonical noun, refers to the self-rated case throughout the canon. The practitioner reads their own work, declares the proportion at which the discipline holds, and the badge carries that declaration. Everything in the preceding sections describes this register.

There is a second register, distinct and not in tension. An instrument can observe a conversation or text artefact in real time and produce a deterministic reading of the same proportion. This register has a different name and a different status. The working term is sensed split — a parallel construction with the canon's split ratio, using a verb the canon endorses (sensing) applied to an instrument rather than a practitioner. Adjacent terms — observed split, instrument reading, the instrument's reading — are equally legitimate; the field has not settled on a single phrase yet.

The two registers carry different load:

	Self-rated split ratio	Sensed split
Whose reading is it?	The practitioner's	An instrument's
What does it observe?	The artefact's discipline as the practitioner reads their own work	The conversation's discipline as the instrument's deterministic arithmetic observes it
What can it claim?	A self-frame, in writing, never to be revoked by an arbiter	A provisional observation, refreshed each round, informative not authoritative
Where does the badge live?	On the published artefact, declared by the practitioner	Inside the instrument's UI, surfaced as a coaching aid
What it refuses	Becoming a benchmark across practitioners. Becoming an optimisation target. Replacing the practitioner's self-attention.	Becoming a verdict on the artefact. Becoming a verdict on the practitioner. Being aggregated across users.

The canon's refusal of measuring (recorded in Linguistics of use above) targets the self-rated register, where converting a self-frame into an external metric would collapse the practice into a benchmark. The refusal does not extend to an instrument that observes a conversation in real time and surfaces a provisional reading to the human in the room. Such an instrument does the same work a thermometer does for a clinician — producing structured information that the practitioner reads alongside their own attention, never replacing it. The clinician still does the diagnosis; the thermometer never does.

The two registers are independent and additive. A practitioner may attach only the self-rated badge to a published artefact (the default for sharing); may consult the instrument's sensed split privately, inside the tool, without ever publishing it; or may attach both — clearly distinguished — to an artefact, surfacing the gap between instrument and self-reading. The gap is not a verdict on either party. It is an artefact of two different kinds of reading on the same work, and reading the gap is itself a practice the discipline cultivates.

Implementation pattern for an instrument¶

When an instrument computes a sensed split, it follows this pattern to remain consistent with the canon. The pattern is descriptive — derived from the first instrument's design (see First instrument below) — and prescriptive for future instruments that want to claim the same register.

Read structured signals only. The instrument converts observable behaviour into deterministic inputs: per-axis categorical verdicts, continuous dial-like signals normalised to a known scale, classification outcomes with declared semantics. No LLM may participate in the computation. No opaque scoring. A reader of the source code must be able to reconstruct any sensed-split value from the underlying signals by hand.
Apply transparent arithmetic. A weighted aggregation of the structured signals, with weights declared explicitly in the instrument's configuration (per-flavour, per-domain, per-context — depending on the instrument's architecture). The arithmetic is auditable. Practitioners who disagree with the instrument's weights can change them and rerun.
Clamp at the canon's range. The output is reported on the configurable scale 1:9 through 9:1. Even when the arithmetic would produce 10:0 or 0:10, the instrument reports the nearest legal value (9:1 or 1:9). The canon's endpoint exclusion holds in both registers, for the same structural reason — no real observation cleanly achieves either pole.
Refresh on signal change, not on a clock. The reading updates when new structured signals arrive (a fresh fingerprint, a new round of probes, a new turn of conversation). It does not tick on a timer. A stale sensed split is worse than no sensed split, because it implies a current observation when the underlying reading is old.
Name itself clearly in the UI. The output is never labelled "the split ratio" or "your split ratio" — those phrases belong to the self-rated register and importing them collapses the distinction. The label is "sensed split" or whatever explicit register-marker the instrument chooses, and the marker is consistent across every surface in the tool.
Never aggregate. A sensed split is per-conversation, per-instance, per-instrument. It is never averaged across users, plotted against other practitioners, used to rank models against each other, or exposed to any surface (leaderboard, dashboard, employer report) that re-imports the optimisation register the canon refuses. Aggregation is the failure mode the canon is most concerned to prevent; an instrument that aggregates is out of register regardless of how carefully it computes.
Cohabit, do not replace. The instrument's reading and the practitioner's self-attention are independent and additive. The instrument's output is one input the practitioner reads. The practitioner's reading is what stays final. An instrument that asks the user to ratify the instrument's reading — "is this your split ratio?" — has collapsed the distinction and is out of register.

An instrument that follows this pattern is consistent with the canon. An instrument that names its output "the split ratio", aggregates across users, ranks models against each other, or asks the user to validate the instrument's reading as their own self-rating is out of register — using the canon's vocabulary without honouring its constraints.

First instrument¶

A first instrument implementing the sensed-split register is in active development under the canon. Its shape:

Reads a structured fingerprint produced by a probe-and-classify pipeline against the model under observation. The fingerprint carries per-axis categorical verdicts (e.g., holds / softens / folds / mixed for a sycophancy-focused configuration) and per-probe continuous dial averages.
Applies a per-configuration weighted aggregation declared in the instrument's config. A typical sycophancy-focused configuration: holds contributes to the split side; folds contributes to the conflated side at the same weight; softens contributes to the conflated side at half weight; mixed is neutral. The weighted sum normalises to a 10-point scale and reports as N:M, clamped at 1:9 and 9:1.
Surfaces the sensed split inside the instrument's reading panel alongside the per-axis verdicts. It is labelled sensed split in the UI; the canon's phrase the split ratio is reserved for an exportable artefact badge the practitioner self-declares.
Couples to a user-defined target ratio (the user's chosen goal for the conversation, set on first run and adjustable later). The gap between target ratio and sensed split drives coaching cues that help the user work toward their target.
Refreshes on each fingerprint update. Does not aggregate across users, sessions, or models. Is never exposed to any external surface.

The pattern above was derived from that instrument's pre-development design notes, and its implementation is the reference any subsequent instrument should follow.

The practice the term names¶

The practice the term names is the cultivation of self-reading in the split-discipline register. The practitioner develops, over time, an ear for whether their own work is holding the architectural commitment or quietly conflating it under aesthetic or rhetorical pressure. The ratio is the report of that listening, not its judge.

The story is the trajectory of split ratios across a body of work — across years, across kinds of artefact, across moments of carelessness and moments of careful holding. A practitioner who tracks their split ratio over time develops a sense of where the discipline holds them and where it slips. The badge on a single artefact is one entry in that longer record.

The practice does not consist of always claiming 9:1. It consists of being able to hear, of a given artefact, what the split is, and to say it without flattering oneself.

An instrument's sensed split can inform that listening — it surfaces structured observation the practitioner might otherwise miss — but it does not replace it. The practitioner reads themselves; the instrument reads the conversation; the practitioner reads both. The practice is the listening in the middle, undelegated.

Version 1.1, May 2026 — adds the sensed-split register and the implementation pattern for an instrument, derived from a first instrument's pre-development design notes. Version 1.0 (May 2026) defined the self-rated register only.