> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tessera.edstratumlabs.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Design principles

> The evidence behind Tessera's design, and the rules it turns into.

Tessera's design comes from a research report covering learning-experience design, LMS usability benchmarks, why Canvas leads the market, evidence on AI tutors and AI-assisted authoring, and learner personas. The full report, with all 72 sources, is on the live site: [research report](https://tessera.edstratumlabs.ai/research.html).

## The thesis

1. **Borrow Canvas's backbone, not its surface.** Canvas leads North American higher education on structure (modules, fast grading, reliability, integrations), not on measured usability: faculty rated it 68.9 on the System Usability Scale, about average. A calmer interface on the same backbone has room to win.
2. **AI is a governed co-author and a hint-first tutor.** In a field experiment, unguarded GPT-4 raised practice scores but lowered exam scores; a tutor that gave hints instead of answers avoided that harm (Bastani et al., *PNAS*, 2025). A carefully scaffolded AI tutor doubled median learning gains over an active-learning class (Kestin et al., *Scientific Reports*, 2025).
3. **Personalize on evidence, not "learning styles."** Style-matching lacks the evidence it needs (Pashler et al., 2008; a 2024 meta-analysis found the required pattern in only 26% of outcomes). Tessera personalizes on goals, time, prior knowledge, language, accessibility, device, and role.
4. **Agents propose, people apply.** Bulk or agentic actions appear as previewable change sets.
5. **Quality and accessibility are checked before publishing,** including AI-generated content.

## The 15 principles

| #  | Principle                                                                                   |
| -- | ------------------------------------------------------------------------------------------- |
| 1  | Modules are the single canonical course spine. AI edits it and never bypasses it.           |
| 2  | The next action is one click from login, and one Today list spans all courses and paths.    |
| 3  | Institution and program templates enforce consistency across courses.                       |
| 4  | Lessons are chunked with time estimates and embedded retrieval checks.                      |
| 5  | The tutor is hint-first, grounded in course sources, and bound by instructor policy.        |
| 6  | AI output is a labeled draft with provenance, a diff, and revert, and a person approves it. |
| 7  | The system explains its adaptations and offers undo.                                        |
| 8  | AI capabilities and limits are disclosed, with no sparkle-only AI cues.                     |
| 9  | Personalization uses evidence-based variables, never learning styles.                       |
| 10 | WCAG 2.2 AA holds across the interface and generated content.                               |
| 11 | Agentic actions are previewable change sets.                                                |
| 12 | A quality and alignment linter runs before publishing.                                      |
| 13 | Interoperability covers LTI 1.3, SIS/HRIS, SCORM, xAPI, cmi5, open APIs, and MCP.           |
| 14 | AI governance is visible: models, data residency, retention, logs, FERPA/GDPR.              |
| 15 | Features are measured against learning outcomes, not engagement alone.                      |

## The tutor modes

Instructors choose a tutor mode for each activity, within limits an administrator sets for each program.

| Mode        | What the tutor does                               |
| ----------- | ------------------------------------------------- |
| **Off**     | No tutor on this activity.                        |
| **Hints**   | Questions and hints only. Never gives answers.    |
| **Explain** | Explains concepts and examples, not item answers. |
| **Open**    | Answers directly. Practice only.                  |

The answer key is never a tutor source. Learners always see which mode is on, who set it, and what their instructor can see.