The Evidence

Abstract

Practicum is a writing design system built on a single claim: writing is where thinking gets done, and anything that shortcuts the writing process shortcuts the thinking. Every architectural decision in the platform — what the AI is permitted to do, how the phases are sequenced, why planning and drafting are spatially separated — can be traced to a body of research on writing cognition that spans nearly forty years.

This paper summarises that research, explains how each finding shaped a specific design decision, and is honest about where the evidence ends and principle begins. It is not a randomised controlled trial of Practicum itself. It is a design rationale: an account of the reasoning behind the system before pilot data exists to confirm or challenge it.

1. The Writing Loop

Practicum structures every writing project through six phases: Read → Pin → Plan → Draft → Review → Edit. This is not a product decision. It is a model of how competent writers actually work — a model described independently by researchers across several decades and traditions.

The loop maps directly onto the most validated framework in writing strategy instruction: Graham and Harris's Self-Regulated Strategy Development (SRSD) POW mnemonic — Pick ideas, Organize notes, Write (Graham & Harris, 2003). SRSD has the highest effect size of any writing intervention in the literature (ES = 0.82; Graham & Perin, 2007). The loop is not a UI convention. It is the structure that the evidence says works.

2. Planning Before Drafting

The finding

Pre-writing activities improve writing quality (Graham & Perin, 2007, ES = 0.32). Structural planning in particular addresses "blank page anxiety" by externalising the question what happens next? before a writer is asked to answer how do I say it?

Kellogg (1988) demonstrated that writers who plan outlines allocate more working memory to sentence construction — because the structural problem is already solved. Planning is not a preliminary to writing; it is part of the cognitive work of writing.

The design decision

Practicum enforces a hard separation between planning and drafting. They are different phases, different screens, and different database columns (Story.Metadata vs Story.Content). The Draft phase does not activate until the Plan is complete. This is not a gamification mechanic. It is a cognitive load decision: the Draft environment is intentionally free of structural prompts so that the writer's working memory is available entirely for prose.

The Bereiter–Scardamalia distinction

Bereiter and Scardamalia (1987) described two modes of writing: knowledge-telling (novice writers write until they run out of things to say) and knowledge-transforming (expert writers plan, wrestle with ideas, then write). The Planner phase is designed to force knowledge-transforming behaviour by requiring the writer to engage with structure before prose becomes available. Every planning prompt is framed as a question, not a suggestion — because the question engages the writer's problem-solving capacity, and the suggestion short-circuits it.

"The hero leaves the known world. What is the point of no return?" — A Practicum planning prompt (question form)

"The hero crosses the threshold by leaving their village." — What the same prompt would say if AI were generating content

The first engages the writer. The second replaces them.

3. The Role of AI — Assistive, Not Generative

The finding

Recent research on AI writing tools raises a consistent warning: over-reliance reduces comprehension, confidence, and independent skill development (Amani & Bisriyah, 2025). Elementary students using AI for creative writing report authorship confusion — uncertainty about whether the work is theirs (LaMear & von Gillern, 2025). Students who use AI to complete writing tasks show reduced self-efficacy over time compared to students who use structured scaffolding to complete the tasks themselves (Kiziltas, 2025).

The design decision

In Practicum's writing applications (StoryFort and Paper Airplane), the AI is constrained to a Socratic role. It asks questions. It does not answer them. It does not generate prose, complete sentences, suggest phrasing, or insert content into the writing area. This constraint is not a UI decision — it is enforced in the prompt layer, where the AI persona's system prompt explicitly prohibits content generation. Removing this constraint would require an intentional code change, not a settings toggle.

The distinction we find useful: a wheelchair ramp is assistive technology. It enables access while the person does the work. An escalator is generative technology — it replaces the effort entirely. Practicum is a ramp.

The Prompt Workbench exception

Prompt Workbench, Practicum's third application, is a professional tool for AI practitioners. Here AI collaboration in authorship is not only acceptable but the point — the audience of the work is an LLM, not a human reader. The same design system applies; the constraint does not. The principle is consistent: AI earns its place relative to the purpose of the work.

4. Privacy as a Design Constraint

The principle

Most privacy failures in software begin before any breach — in systems that collect more data than they need, concentrating sensitive information that becomes a target. The most effective safeguard is not a stronger lock on a vault; it is a vault that contains almost nothing worth stealing.

Practicum collects the minimum personal data needed to operate. It collects no personally identifiable information on children. This is not a compliance posture. It is an architectural decision made before the first line of code was written, grounded in the conviction — held personally, not theoretically — that the cost of a breach lands on real families.

What this means in practice

Student accounts within StoryFort's classroom management system are identified by a class code and a student-chosen display name. No email address. No date of birth beyond a school-year cohort for age-gating. No usage tracking beyond what is necessary to resume a writing session. The data model was designed so that a hypothetical breach of the student database would expose nothing a bad actor could use.

5. Where the Evidence Ends

This paper has described a design rationale, not a trial. Practicum has not yet been studied in a classroom. The research cited here supports the principles behind the design — it does not prove that this specific implementation produces the outcomes the research predicts.

That work is next. Practicum is approaching its first classroom pilot. The goal is to instrument the platform to measure what the research says matters: whether students complete the planning phase before drafting, whether the Socratic AI prompts produce more independent idea generation than open-ended drafting support, and whether students who use the system report stronger authorship ownership over time.

The evidence summarised here is the foundation. The pilot is the test.