SDC in brief

PlateLanguage describes; judgement decides. The gap is the architecture.

One sentence

Split-Domain Cognition is a principle that keeps language work (descriptive, generative, interpretive) and judgement work (evaluative, verdictive, categorical) structurally apart, so that both serve the intent of the person doing the work; fusing them in a single call reliably produces the same failures — unauditable verdicts, descriptions that smuggle evaluation, rubrics that cannot be taught, drift under pressure, sycophancy in machines.

One paragraph

Split-Domain Cognition is the observation that, whenever a single agent — a person, a model, an institution, a form — is asked to do two unlike kinds of thinking in the same breath, a predictable pattern of failure appears. The first kind of thinking is descriptive: it attends to what is there. The second is verdictive: it closes on an outcome. The two can look similar from outside and feel like one continuous act from inside. They are not one continuous act. Their criteria are different in kind: language work goes right by coverage-under-interpretation, judgement work by reproducibility-under-rule. Optimised in the same channel, they contaminate each other. SDC is the architecture that refuses the contamination — keep the two domains structurally apart, let each be answerable to the criteria proper to itself. Both serve that intent, and neither domain sets it.

What the split serves

Language work and judgement work serve an intent: what the work is for, held by the person doing it, even when somebody else owns the work. Ownership decides that the work gets done; how it is done, and what it is for beyond the job, come from the one doing it. The split keeps the two domains from borrowing each other's authority; it does not decide what they are for. A split kept perfectly can still serve somebody else's purpose, so the first question is whose intent the work serves, and the second is how to keep language and judgement apart in its service. The intent is never handed to a model or written into a rule.

Who holds the language

The split does not give the language side to a model by default. A model does language work only where that has been negotiated and set up for that work; otherwise the language is the people's: a model at most reads an answer and points to the relevant part, and code decides. That negotiation is only honest where what the person says they want and what they feel they want match; the sentence that states the intent is always theirs.

The AI-pipeline instance

In software, SDC is most commonly expressed as a three-stage pipeline.

1. Stage 1 — Language work. A language model reads unstructured input (concept statement, transcript, sketch description, menu, brief) and produces structured signals: classifications, embeddings, extracted features.
2. Stage 2 — Judgement work. Deterministic code applies thresholds, priorities, and rules to those signals and produces a verdict: a score, a ranking, a surfacing decision.
3. Stage 3 — Language work (again), only where negotiated. A language model narrates the verdict in readable prose, if that has been negotiated and set up for that work. Otherwise there is no third stage, and the people using the tool write the language.

The three stages are held apart on principle. Language models are good at pattern recognition across language but unreliable at verdicts that must be consistent, auditable, and reproducible. Code is good at verdicts but cannot generate fluent narration without overwhelming the user. Each stage does what it is good at. None borrows the criteria of another.

Why this is a principle, not only a pipeline

The pipeline is the most compact expression of SDC. It is not the whole of SDC. The same move shows up in teaching (rubric-bound critique separated from generative making), curation (description separated from curatorial thesis), animal ethics (rights-grounded judgement separated from welfare-language description), editorial work, and many other places.

The AI instance makes the move precise because software forces precision. But the move is older than the software and will outlast any particular implementation. See the position essay for the longer argument. ¹ The degree to which any single artefact holds the discipline is named, self-rated, by The Split Ratio.

Where this sits — the lineage, and what is new

Reading qualitative material and producing structured, numerical signals from it is a named, mature practice, and SDC names its lineage rather than reinventing it. In mixed-methods research the act is called quantitizing (the conversion of qualitative data into quantitative form); in the social sciences, quantitative content analysis, with its long discipline of intercoder reliability (Krippendorff's alpha, Cohen's kappa); in computational social science, text-as-data, governed by the rule that automated methods are no substitute for close reading and demand problem-specific validation. Confining a language model to a coding role is, today, an active and crowded field. SDC's first stage is, in this light, a quantitizing instrument.

What SDC adds is narrow and specific.

1. The field quantitizes and then judges in the same breath — the same coder, or model, both codes and scores. SDC holds the language work structurally apart from the judgement and forbids the language layer from issuing verdicts.
2. The established discipline validates automated coding against human agreement as ground truth. SDC's target is auditability against an explicit rule, not agreement with a human judge — so a method can agree only modestly with human coders and still be doing exactly what SDC intends.
3. Quantitizing lives in social science; SDC carries the same split across many domains — pedagogy, curation, ethics, and the shape of a practice.

The novelty is the discipline of the split, not the act of conversion.

When does SDC apply?

All of these must hold:

1. Two kinds of cognition are in play. Generative/interpretive and evaluative/verdictive activity, both present.
2. They are at risk of collapsing. In current practice the two are being done in the same call — same person, model, form, moment.
3. The collapse produces a measurable cost. Trust eroding, decisions indefensible, learners cannot learn the criteria, stakeholders cannot re-enter the argument.
4. The judgement can be codified. The verdict is one an explicit rule can carry — a criterion or threshold a person can write, read, apply, and contest (whether applied by a script or by a human bound to the written standard). This is a line drawn inside a judgement, not a gate on whole subjects: most judgements are mixtures, and SDC takes the codifiable part while the irreducibly interpretive remainder is named as residue. Only a verdict reducible to nothing but a contested reading falls wholly outside.

If these hold, SDC is appropriate. If fewer, it is probably not.

When does SDC not apply?

• Generative-only work (a painting, a poem, an improvisation).
• Judgement-only work with no language component (a tax calculation, a chess move).
• Domains where the fusion is the point (oratory, advertising, evangelism).
• Decisions where speed matters more than auditability (emergency triage).
• Irreducibly interpretive judgement, where no formal rule can carry the verdict. Sedition is the type case — kept interpretive rather than codified, which is what makes it abusable. Here SDC's role is only diagnostic: to mark where an interpretive judgement is dressed as a codifiable rule.

A principle with a clear scope is a principle with teeth. SDC declines to be everything.

---

1. The position essay is at A Principle, Not a Pattern. The long-form articulation is published as Split-Domain Cognition (PDF). The full set of canonical documents — long form, position essay, operational Process, derivation protocol, and the canonical variants — is listed on the Canon page. ↩