No human participants. No measured comprehension or retention gain. This model tests the sensitivity of planned reading paths to assumed reading, diagram-decoding, navigation and interaction costs. It does not model people as if they had actually visited the page.
The original purpose is A: recognise the existing coordination burden; B: explain the evidence → forecast → policy mechanism; C: identify the bounded assistance and how its value would be tested. The first impression must also answer “why bother?”
Text exposure is measured from the rendered article. Diagram labels count as words. Reading speeds, figure-decoding times and navigation operations are declared design assumptions. More diagrams can aid a mental model or make decoding harder; their time is never free.
T = reading + figure decoding + navigation + control use
A lower modelled time is not evidence of correct understanding. No result is called an attention improvement or a probability that a person will stay. There is no v6-versus-v7 speed claim because the restored scope changes the reader's task.
Task: Explain the problem, steps and outcome without opening the maths.
Likely friction to inspect: Mistake a proposed result for a working personal service.
Design response: Outcome labelled illustrative next to the concrete result; status above title.
Task: Follow one evening and save a task-specific implementation brief.
Likely friction to inspect: An interesting vision with nothing actionable to hand an agent.
Design response: Brief exports the chosen scenario, costs, outcomes and prohibited actions.
Task: Decide whether the token and human-workload assumptions are worth testing.
Likely friction to inspect: Confuse a low token bill with a profitable or equal-quality system.
Design response: Separate retained events, model routing, useful outcomes, upkeep and setup cost.
Task: Recover the system shape and use a familiar scenario without replaying animation.
Likely friction to inspect: Small screens make spatial connections hard; animation forces waiting.
Design response: Responsive map, visible labels and direct stage controls; no essential hidden content.
Task: Inspect forecast updates, test design, safety boundaries and old counterexamples.
Likely friction to inspect: Synthetic, mathematical and personal claims blend together.
Design response: Dedicated companion and explicit evidence levels, preserved source results.
Task: Understand A/B/C without holding one distant paragraph in working memory.
Likely friction to inspect: Labels become another unfamiliar notation system.
Design response: One concrete evening, nearby source/output labels and a plain-language baseline.
18,000 parameter draws; 5th–95th ranges are sensitivity ranges, not confidence intervals or forecasts of reader performance.
| Reading path | Diagram assumption | Median | 5th–95th range |
|---|---|---|---|
| goal_first | familiar_map | 46 | 36–61 |
| goal_first | ordinary_decoding | 57 | 44–73 |
| goal_first | difficult_diagram | 94 | 68–118 |
| busy_builder | familiar_map | 178 | 146–220 |
| busy_builder | ordinary_decoding | 208 | 169–258 |
| busy_builder | difficult_diagram | 312 | 240–390 |
| cost_sceptic | familiar_map | 215 | 173–282 |
| cost_sceptic | ordinary_decoding | 246 | 196–316 |
| cost_sceptic | difficult_diagram | 359 | 269–449 |
| mobile_return | familiar_map | 67 | 55–88 |
| mobile_return | ordinary_decoding | 79 | 63–100 |
| mobile_return | difficult_diagram | 116 | 88–143 |
| method_reader | familiar_map | 801 | 644–1073 |
| method_reader | ordinary_decoding | 847 | 689–1121 |
| method_reader | difficult_diagram | 1029 | 834–1318 |
| low_reserve | familiar_map | 69 | 55–95 |
| low_reserve | ordinary_decoding | 79 | 62–107 |
| low_reserve | difficult_diagram | 115 | 87–150 |
Before exposing the article, assign a relevant task. After the first reading opportunity, ask for an unprompted explanation of the problem, the mechanism and the proposed result. Score the distinction between a forecast, a goal and authority to act. Then transfer to a new household or coding scenario.
Measure first-correct explanation, errors, transfer, delayed recall, voluntary continuation and perceived effort separately. Record non-completion rather than dropping it from analysis. Screen readers, mobile use, reduced motion and interrupted reading need actual participants; synthetic personas do not stand in for them.
Predefine acceptance criteria and the comparison task before collecting results. Reading time alone is not the target: a longer explanation that is correct may outperform a fast but misleading skim.
The previous v6 model, protocol and its mixed findings are retained in the companion ZIP. The new design does not overwrite those unfavourable sensitivity cases.