Choice Architecture Is a Handoff Problem

A syllabus for a category theorist — and why the seam is the unit of design

by N. P. Clarke — September 2026, for a friend who bonded with me over category theory, is stuck on gluing, thinks in Dirac and Pauli, and had never heard of choice architecture

Why a sales method needs this subject

SparkTorch's claim has always been that the hero is the Handoff — the seam where one person's trust becomes another's. A Handoff is a re-presentation: the same person, the same offer, the same relationship history, presented to a new party in a new context. Whether trust survives that change of presentation is the whole question. Choice architecture is the discipline that studies exactly that question — what happens to a choice when only its presentation changes — and it has a fifty-year evidence base that the crew should know and the Spark should feel.

The field's front door is behavioral economics, which is why mathematicians and engineers walk past it. Its back door opens onto two things a category theorist is already holding: non-commuting observables and the failure to glue. This syllabus enters through the back door.

Move 0 — What the field says it is

Choice architecture (Thaler & Sunstein, Nudge, 2008) is the design of the environment in which a choice is made: the default, the order, the framing, the grouping, the description. The founding fact is that choices are not invariant under re-presentation of the same option set. Flip the default and organ-donor consent moves from roughly 15% to roughly 90% between otherwise similar countries (Johnson & Goldstein, Science, 2003). Say "90% survive" instead of "10% die" and preferences move (Tversky & Kahneman, Science, 1981). Johnson et al., "Beyond nudges: tools of a choice architecture" (Marketing Letters, 2012), is the catalogue.

In SparkTorch terms: Gate 0 is a choice architecture. Deciding what not to present — the relevance wedge — is the first architectural act. Every buyer meets an option set that someone already shaped.

Move 1 — Presentation is not the object

Rational choice theory assumes description invariance: the choice function is defined on the option set, so it factors through the quotient by "same options, different words." Choice architecture is the experimental discovery that it does not factor. Choice is a functor of the presentation — the chart — not of the underlying set, and the nudge literature is a catalogue of the ways naturality fails across charts.

The cleanest case is the default. A choice architecture is a pointed presentation; the default is the basepoint; the effect size is the price of leaving it. The dignity line in the method — never spray the buyer — is a rule about basepoints: a default that serves the buyer's own interest is architecture; a default that exploits the cost of leaving it is spray.

Move 2 — The orders don't commute

Ask two attitude questions in order A→B and B→A and the answer distributions differ — a nuisance in survey methodology for decades. Wang and Busemeyer modelled the two questions as non-commuting projectors on a belief state and derived a parameter-free prediction, the QQ equality (Wang & Busemeyer 2013). They tested it on 70 nationally representative surveys and it held (Wang et al., PNAS 2014). Pauli speculated in 1950 that complementarity had consequences outside physics; the 2013 paper quotes him.

The method's phases — Spark, Handoff, Torch, Relay — are ordered, and the order is not decoration. Whether the buyer meets the Spark before the artifact or the artifact before the Spark changes what the buyer believes afterward; that is an order effect, and the formalism that predicted the survey regularity is the one to reach for.

Now the Dirac-and-Pauli temperament should push back, and the literature has done it already: classical repeat-choice models reproduce the QQ equality (Kellen et al. 2017); Dzhafarov shows a QQ-satisfying system is, in his sense, noncontextual — the equality is evidence of order-dependence, not of quantum structure (Dzhafarov et al. 2017); and the projective model cannot jointly hold order effects and response replicability without moving to quantum instruments (Ozawa & Khrennikov 2020). The honest state: the mathematics transfers; the metaphysics does not.

Move 3 — Gluing: seam integrity is a sheaf condition

Abramsky and Brandenburger gave contextuality a single sheaf-theoretic definition: an empirical model is a presheaf of local distributions over a cover of contexts, and contextuality is exactly the obstruction to a global section — local data that are pairwise consistent and yet do not glue (Abramsky & Brandenburger 2011). The same structure appears in database theory and logical paradox, with cohomological witnesses (Abramsky, Barbosa & Mansfield 2015).

Read a Handoff through it. The Spark's context and the Torch's context are two charts on the same buyer. Within each, the buyer's trust is locally coherent. Seam integrity — the method's own metric — is the question of whether those local sections glue into a global one. A Handoff that fails is not a bad conversation; it is a buyer whose trust in the Spark's frame and trust in the Torch's frame do not extend to a single section. Preference reversals in the lab are the same object: locally rational, globally unglued. A "rational agent" in the textbook sense is a global section, and a nudge is a change of cover.

Two recent pieces make the gluing explicit — a gluing functor on Boolean contexts whose failure to stay Boolean is equivalent to the absence of global sections (Gunji et al. 2026), and Dzhafarov's translation of the sheaf approach into Contextuality-by-Default so it applies to behavioral data where even the marginals disagree across contexts (Dzhafarov 2019). If you are stuck on gluing in the abstract, here is a place where the failure to glue is a measured fact about people, with a cohomology class attached — and a sales method that has been trying to measure it with a clipboard.

Move 4 — The evidence, said honestly

Does engineered choice architecture work? The big meta-analysis found a small-to-medium effect, d ≈ 0.45, across 200+ studies and 2.1 million participants — with "moderate publication bias" (Mertens et al., PNAS 2021). Three replies argued the bias was the headline: after robust Bayesian correction, no evidence for an overall effect remains (Maier et al. 2022); the effect distribution is left-truncated at zero, the signature of suppressed nulls (Bakdash & Marusich 2022); and the conditions for success are barely identified (Szaszi et al. 2022). A 2025 second-order meta-analysis over 1,638 studies and about 30 million people: d = 0.27, dropping to 0.004 after bias adjustment (Hu et al. 2025).

So separate two claims the field blurs. The invariance failures — framing, defaults, order — are robust facts; Moves 1–3 stand on them. The policy efficacy of engineered nudges is contested and probably small on average, with defaults and decision-structure interventions surviving best. That split is the method's position exactly: presentation shapes trust — and no amount of presentation substitutes for a human who has earned it. The evidence says the manipulation was oversold; the seam was not.

Move 5 — Language and notation are choice architectures

Framing is linguistic: "90% survive" and "10% die" are one option set under two grammatical presentations, and the enriched-category treatment of language (meaning from compositional structure) is the right formal home for "which re-descriptions are not identity." The method's rule to build artifacts in the buyer's domain language and the expert's voice is a framing rule with a reason: fluency is the trust signal because the buyer's own grammar is the chart in which their trust is already a section.

And a notation is a choice architecture for a composer. Staff notation makes twelve-tone equal temperament the basepoint — cheap to write, expensive to leave — which is why microtonal and electroacoustic composers keep inventing notations to move the default. Chinese percussion notation defaults to timbre and stroke instead of pitch, and a different music becomes cheap. A notation's obligatory categories are its type system; its defaults are its nudges. Every grammar has a basepoint.

What the crew does, and what it never does

The crew keeps the presheaf. Machines can carry the frames, log which presentation each party saw, compute whether the local sections are consistent, and flag the obstruction — that is the ledger work of the Handoff, and it is worth a great deal because most failed seams fail from forgetting what was promised in which frame. What the crew cannot do is post the collateral. It has no relationship history to pledge, so when it re-presents a buyer to a new context, nothing of its own is at stake, and a Handoff with nothing at stake is not a Handoff. The crew computes the obstruction; the humans glue. Empower, never replace — here as a theorem about sections, not a slogan.

The exercise

Take a classic preference-reversal dataset (Lichtenstein & Slovic's P-bets versus $-bets). Model the frames as a cover and the choice distributions as a presheaf, and ask whether a global section exists. If not, compute the cohomological obstruction. Then take a QQ-satisfying survey pair and run the Contextuality-by-Default test — you should find it noncontextual, which is the surprising result: order effects and contextuality are different failures. Write down the difference in sheaf language. Then do the same for a Handoff log: two frames, one buyer, pledged trust on both sides of the seam. If it glues, seam integrity is a number. If it doesn't, you have found where the method is standing on a hole.

That seam is empty in the literature. It suits a category theorist.

Reading, in order

  1. Thaler & Sunstein, Nudge (2008; final ed. 2021), chs. 1–5 · Johnson et al., "Beyond nudges: tools of a choice architecture," Marketing Letters 23 (2012) · Johnson & Goldstein, "Do defaults save lives?" Science 302 (2003) · Tversky & Kahneman, "The framing of decisions and the psychology of choice," Science 211 (1981).
  2. Wang & Busemeyer 2013 · Wang et al., PNAS 2014 · Yearsley & Busemeyer, tutorial 2016 · pushback: Kellen et al. 2017, Dzhafarov et al. 2017 · Busemeyer & Bruza, Quantum Models of Cognition and Decision (Cambridge, 2012).
  3. Abramsky & Brandenburger 2011 · Abramsky, Barbosa & Mansfield 2015 · Dzhafarov 2019, the CbD view of the sheaf approach · Gunji et al. 2026, contextuality as a left adjoint.
  4. Mertens et al. 2021 · Maier et al. 2022 · Szaszi et al. 2022 · Bakdash & Marusich 2022 · Hu et al. 2025.
  5. Tai-Danae Bradley, "Where does meaning live in a sentence? Math might tell us," Quanta (2025) · Hertwig & Grüne-Yanoff, "Nudging and boosting," Perspectives on Psychological Science 12 (2017) — the boost is the human rail: raise competence, don't steer.

Choice, made legible at the seam.