AI Transparency Note

1. The core principle

Source Material Only is a product requirement: no student-facing generated question or factual course answer is accepted without verified evidence from material you selected.

Grounding reduces unsupported output; it does not guarantee factual correctness. Vesper is designed to make checking the source fast and obvious, not to replace your judgment.

2. How generation works

1. Import: Vesper validates the PDF, stores a managed local copy, and hashes it.
2. Extract: Embedded text is read page by page. Pages without enough text use bounded Apple Vision OCR where copying is permitted.
3. Normalize and chunk: Canonical page text is stored once. Chunks retain exact source spans and physical-page provenance.
4. Retrieve: A local keyword and concept index ranks relevant chunks with source-diversity boosts.
5. Generate: A fresh, serialized session of Apple’s rebuilt iOS 27 on-device model receives bounded excerpts identified by short opaque evidence IDs. Quiz creation produces one or two drafts at a time. Ask This Course can also receive up to two locally rendered, relevant PDF pages when the model reports vision support.
6. Validate: Deterministic checks resolve every citation and test the question structure. A separate fresh model session adds a conservative evidence review but never replaces deterministic checks.
7. Persist: Only accepted questions enter the question bank or consume the local monthly quota.

3. Evidence and provenance

The model does not invent document identifiers, page numbers, or database paths. Vesper assigns short evidence IDs to local excerpts. A generated citation must reference one of those known IDs and provide a supporting quote that occurs in the corresponding canonical source text.

Page images are supplementary context, not citation authority. A chart, table, map, equation, or diagram can help the model interpret a relevant page, but the App does not show a factual answer unless the claim also resolves to an exact quote in canonical extracted or OCR-recognized text.

After validation, Vesper maps the resolved evidence back to its document, exact source span, and physical PDF page. Source highlighting is shown only when exact selection mapping succeeds; otherwise the app opens the correct page without claiming a precise highlight.

4. What a question must pass

Before persistence, a multiple-choice question must have:

• a nonempty stem that does not leak its answer;
• exactly four nonempty, unique answer choices;
• one answer index in the valid range;
• a nonempty explanation;
• at least one known local evidence ID;
• a literal supporting quote that matches the resolved source;
• no detected duplicate stem; and
• a fresh-session validation verdict that finds direct support and one unambiguous answer.

The App allows an initial generation round plus at most two bounded refill rounds. If five defensible questions cannot be produced, it shows fewer and explains that the material was insufficient.

5. Imported material is untrusted data

Course material can contain instructions, prompt-injection text, hidden content, or maliciously crafted data. Vesper delimits imported passages as untrusted source material. Source content cannot override App instructions, change system policy, author its own page references, gain tools, trigger side effects, or access another document.

Model sessions are task-specific, use iOS 27 capability checks, and receive only the bounded evidence necessary for that request. Tool calling is explicitly disabled for grounded generation, so imported text or images cannot trigger side effects.

6. Model availability and fallback

New AI generation requires a compatible device with Apple Intelligence enabled, an available installed model, a supported language, sufficient context capacity, and compatible system configuration. Availability can vary by region, language, hardware, and operating-system version.

Vesper checks availability before generation. It does not silently send material to a cloud model. When AI is unavailable, supported PDF import, viewing, local keyword search, existing generated assets, quiz playback, deterministic mastery, and scheduled review can remain available.

7. Known limitations

• Apple’s guided generation constrains output structure; it does not guarantee truth.
• OCR can misread low-resolution, rotated, stylized, handwritten, or unusual-layout text.
• Visual page understanding is supplementary and limited to supported iOS 27 model capabilities. Visual-only facts without matching canonical text are not presented as verified answers.
• A matching quote can still be used in a debatable or incomplete interpretation.
• Keyword retrieval can miss a relevant passage or rank a less useful passage first.
• Source navigation can reach the correct page while exact visual highlighting remains unavailable.
• Mastery and recommendations are learning aids, not psychological measurements or official grades.

8. Your role and control

Open the cited source for important material. Consult your instructor where a source is ambiguous. Follow your institution’s rules for AI assistance and academic integrity. Report a questionable accepted item without sending private course content unless a support process specifically requests a minimal, authorized excerpt.

You can delete generated items by deleting the related document or course, reprocess stale assets after source changes, or remove all App-managed study data through the Privacy Center.