Users finish the task. They still cannot explain what happened.

People send money, book doctors, renew licenses — and leave the screen unsure of what just happened. The standard explanation is cognitive overload. But the user who finishes without errors and still cannot say whether the money has left the account was not overloaded. They were not informed.

This is a live international experiment that makes that gap visible. I designed the construct (cognitive distance), operationalised it into four dimensions, pre-registered five hypotheses, and shipped the platform collecting data today. The conceptual paper is under peer review.

The result is a research instrument for questions performance metrics have ignored for thirty years — and a demonstration of taking a construct from theory to instrument to data.

Cognitive load theory (Sweller, 1988) gave HCI a vocabulary for capacity: working memory is finite, and tasks that exceed it fail. Decades of design advice followed — simplify, reduce, declutter.

Yet the most consequential usability failures persist in systems that have already lowered load. A user completes a transfer in three taps, sees the confirmation, and still cannot say whether the money has left. The interface was not too heavy. It was too opaque — the user executed system logic they could not parse.

That is the difference this study operationalises. Cognitive load measures how much effort a task demands. Cognitive distance measures whether the user can explain, predict, and transfer the system's logic into their own reasoning. Load is about capacity. Distance is about alignment.

Usability research relies on time-on-task, completion rate, and error frequency. ISO 9241-11 codified them; Nielsen operationalised them; most product teams use them weekly. They are robust — and none of them answers: did the user understand what just happened?

A high completion rate can sit on widespread misunderstanding. A low error rate can co-exist with users who cannot say whether their task succeeded. The proxies register success while interpretation collapses.

The cost falls hardest on users with the least interpretive surplus: people with cognitive disabilities, neurodivergent users, older adults, anyone working under time pressure or in a second language. For them, opacity translates directly into exclusion, dependence, and repeated error.

Moving cognitive distance from slogan to construct requires treating it as a multidimensional variable. Four observable dimensions — drawn from cognitive ergonomics and HCI — triangulate the construct. Each is independently measurable. Together they form a Cognitive Distance Index (CDI) portable across studies, product lines, and design reviews.

From score to action: four diagnostic bands

The CDI is diagnostic, not comparative. Each band carries a recommended response: Preferred (maintain), Watch, Action needed, and Critical (governance-level redesign).

The design is deliberately austere. Interface language register is the only variable that moves. Every click, every field, every outcome stays identical. If the arms diverge on completion time or effort, the manipulation failed — and manipulation checks guard exactly that.

Between-subjects, not within. Each participant sees one register across all three modules. Once someone decodes an institutional phrase, they carry that decoding forward — between-subjects is the only design that keeps the register stable.

Balanced, not random. Silent server-side assignment picks whichever arm has fewer starters (ties break randomly). Module order is balanced across six rotations of [banking, medical, license]. Device-local duplicate detection prevents re-entry.

Three domains, one protocol. Online banking (transfer), medical appointments (specialist booking), and driver license renewal (pay, review, mail). Each module follows the same arc: task intro, prototype flow, prediction MCQs, result screen, comprehension MCQs, post-task scales.

Manipulation checks guard isolation. If the arms diverge on ease, effort, or completion time, something other than language moved. The platform flags this in real time on the researcher panel. Failed checks trigger sensitivity analysis.

A good research design means nothing without a platform participants trust, on the devices they use, in the languages they read. Building it end-to-end — front-end, back-end, deployment, researcher panel, walkthrough mode — was the bulk of the work. Six decisions mattered most.

Another researcher could rebuild the protocol from public documentation alone. Module rules live in plain Markdown (RULES-BANKING, RULES-MEDICAL, RULES-LICENSE). Business rules, Version A/B result copy, and scoring keys for every question are versioned alongside the code.

Anonymisation is aggressive by design. No name, email, IP address, device identifier, or full user-agent in the research dataset. No participant cookies. The researcher panel uses an httpOnly session cookie scoped to staff. This shaped the consent flow, welcome copy, assignment route, and deletion API — not a compliance afterthought.

The construct points at a recurring design pattern that produces cognitive distance. The conceptual paper introduces externalized processing: systems that specify valid input conditions without providing internal means to meet them. Format restrictions, size limits, naming rules — each becomes a user-side preprocessing task with an interpretive burden on top.

The diagram below contrasts the pattern that pushes work outward (letting interpretation collapse) with the pattern that keeps it in-system (keeping the causal chain legible). Every module embeds a version of this trade-off, which the comprehension and prediction questions are designed to detect.

The platform reads as a calm research instrument — not a survey app, not a marketing flow. The question dock keeps results visible while participants answer. Result screens name the next actor and timeline in plain language (Version A) or in the system's own voice (Version B).

Data collection is active. The observations below are descriptive from early cohorts — directional, not headline findings. Inferential analysis follows manuscript submission. The panel surfaces every chart with that caveat inline.

If the hypotheses are supported, the study shifts what counts as a usability problem. Three implications travel with the construct.

1. Accessibility moves from accommodation to prevention. Neurodivergent HCI research shows systems designed around neurotypical defaults impose invisible access work. Cognitive distance makes that work visible and measurable in design review.
2. AI-mediated interaction becomes diagnosable. Automation, adaptive logic, and AI-generated outputs produce what Sarter, Woods, and Billings called "automation surprise" — usable in the moment, opaque in rationale. CDI offers a mediating variable for trust, fairness, and reliance.
3. Evaluation redefines what counts as evidence. Beyond whether users succeeded: can they explain the response, predict the next state, and identify the causal structure connecting action to outcome? That is the evidence base for inclusive, accountable design.

Next.js 16 (App Router), React 19, Tailwind CSS 4. Supabase (Postgres + Row Level Security) for participants, module_runs, events, answers, and open_scores. Vercel deployment with /api/health check. No PII in the research dataset; no participant cookies; httpOnly session cookie for researcher access only.

The platform is live, collecting data internationally. The conceptual paper is under peer review. The framework extends to other services, locales, and types of opacity — automation, AI, data dashboards.

This case study documents the work of taking a construct from intuition to instrument: a research design that survives review, a platform participants trust, and a measurement practice honest enough to publish.