Three personas came out of the research. Their job wasn’t decoration — it was to force a prioritization.

Daniel has the app and uses it like a weather forecast. He doesn’t know it’s a two-way tool. Told that it is, he assumes the reporting side is for scientists, not him. He’s the biggest group by far, and he’s the discovery leak.

Selma took an avalanche course last season and feels a real duty to contribute. But every submission feels like a high-stakes test she might fail, and with no feedback, she never learns she’s doing fine. She reports a fraction of what she could. She’s the activation leak — and the one most at risk of churning out entirely.

Arian is a paid expert. He knows exactly what he wants to say and finds the form’s rigid structure gets in his way. He’s built workarounds — handwriting notes to type up later, buying a stylus so he can report with gloves on. He’s tolerant of a bad form because he’s embedded and on payroll. He’s the leak I chose not to chase.

This ran alongside a full-time internship, in a domain I didn’t know, with no team. AI tools filled parts of that gap — but the division of labor is the whole point.

ChatGPT held the project’s working memory. I kept all the context in one place and used it as a standing sparring partner: stress-testing my domain assumptions, comparing candidate directions while scoping, and — because sessions were short and scattered — reminding me where I’d left off when I hadn’t left myself notes. Claude did the heavy synthesis, including turning the hand-catalogued form inventory into the interaction cost analysis. Gemini wrote prototype code. Image models made visual assets.

What AI didn’t do: talk to users, decide what mattered, or notice what was missing. And the real risk was never hallucination — it was plausible overclaiming. Drafts kept drifting toward conclusions the evidence didn’t support, and a language model will amplify that drift eagerly. The most useful habit I built was interrogating every claim with one question: is this a finding, an applied rationale, or a hypothesis? Most of the honesty caveats in this case study exist because of that question.

I put every candidate intervention on a 2×2 of effort versus impact, then grouped related changes into clusters.

A lot of the high-impact work turned out to be low-effort: swapping dropdowns for chips, replacing keyboard number entry with steppers, hiding fields a given user doesn’t need. Cheap to build, cheap to test, and they cut real friction. That shaped the whole prioritization — instead of betting on complex new features, I focused on foundational fixes I could validate quickly against the large dormant user base, leaving a clean foundation to iterate on once there’s real usage data.

The interaction cost model started as a communication problem: I needed to compare the current and redesigned forms with NVE in terms more concrete than “it feels simpler.”

The method: I catalogued every form field in both versions by hand — input type, conditional logic, repetition patterns. Each input type got a weight in interaction cost points, informed by walkthrough sessions where active observers narrated which interactions felt effortful, a small branched survey of experienced reporters rating per-section effort, and the interview data. A chip costs 1p. A keyboard numeric input, 3p. Extended free-text composition, up to 16p. A location pin on a map, 10p. Summed across realistic scenarios, a typical comprehensive observation drops from ~669p to ~463p — the 31% — and the effort gap between a quick and a comprehensive report narrows from 56× to 26×, which matters for anyone weighing the climb from casual to serious contributor.

Then I asked whether the model was any good, and fell down a hole. It turns out HCI has been trying to measure interaction effort rigorously for fifty years, and the field is split down the middle on how.

One tradition predicts effort before anyone touches the design. Card, Moran and Newell’s GOMS and its stripped-down cousin the Keystroke-Level Model decompose a task into primitive operators — keystrokes, pointing, mental preparation — each with an empirically measured time constant. Point at a target and Fitts’s Law tells you, to the millisecond, how long the movement takes as a function of distance and target size. These models are the gold standard for a reason: cardinal, validated across decades, runnable on a wireframe. But they buy that rigor by modelling a narrow slice of reality. KLM assumes a skilled user performing without errors, and collapses all of thinking into a single ~1.35-second “mental operator.” Fitts covers the finger, not the mind.

That’s fatal for a form like RegObs, because its real cost was never in the tapping. It’s in deciding what you observed and which field it belongs in — translating a hunch about the snowpack into the system’s categories. That’s Norman’s gulf of execution, and it’s exactly the cognitive load the predictive models wave away with one constant. Measure RegObs with KLM and you’d conclude the hard part is cheap.

The other tradition measures effort after the fact, from the person who felt it. NASA-TLX asks users to rate a task across mental demand, physical demand, temporal demand, performance, effort and frustration — capturing precisely the cognitive and affective weight the objective models miss. The catch is symmetrical: it needs real people doing real tasks, so you can’t run it on a design that doesn’t exist yet. Nielsen’s “interaction cost,” the practitioner phrase I’d borrowed the name from, sits deliberately loose — the sum of mental and physical effort, no unit, no formula.

My model is a hybrid, and its flaw lives exactly on the seam. I took ordinal data — perceived-effort rankings from the experiential tradition — and did arithmetic on it, summing and ratioing as if it came from the predictive one. Stevens’ measurement theory is blunt about this: ordinal scales support order, not addition. “3p is harder than 1p” is a legitimate ordinal claim; “3p is three times 1p” and “463 is 31% less than 669” quietly promote it to a ratio scale it was never built on. So the 31% is directionally defensible and metrically fictional at the same time. The model also predicts no completion times, has no statistical validation, and conflates motor with cognitive effort — the very conflation the two canonical traditions exist to keep apart.

There’s a blunter truth under all this, and it took me a while to let myself say it. The form was obviously broken — made once, forgotten, never revisited. You don’t need a validated instrument to measure a thing with a hole in it, and reaching for one risks spurious precision on an object that never earned it. The trap in “obvious,” though, is the question it invites: if it was so plain, what did I actually do? The answer is that obvious and trivial aren’t the same. The form had been normalized by the people closest to it — its brokenness was masked by neglect, not by subtlety. Re-seeing what everyone had stopped seeing, and quantifying it just enough to make it undeniable, was the work. The obviousness was a finding, not a given.

So why keep it? Because its actual job was never prediction. It was comparison and communication — persuasion-grade rigor for a problem that was already plain — and for that it earned its keep twice over. It triaged design decisions, pointing straight at the temperature and density layer sections, whose per-entry costs compounded worst across repeated entries. And it gave stakeholders a single number to argue over, which prose about “friction” never achieves. The lesson isn’t “the model was wrong.” It’s that rigor should match stakes: effort measurement forces a choice of tradition, and the honest move is to pick the one your question actually needs — or, if you’re going to straddle them like I did, to say so out loud rather than let a clean-looking percentage imply a precision it doesn’t have.