How GraspN works, and what it does with your reading.

Documentation

GraspN sits on top of what you were already reading. It turns a page, a PDF, or a pasted passage into questions grounded in that source, and keeps the result as a model of your understanding rather than as a saved quiz.

The pipeline is three stages with a typed contract between each one, so any of them can improve without the others changing.

1. STAGE 01CAPTURE

   Material is versioned and split into ordered source blocks. Block numbering survives re-imports, which is what lets a citation keep pointing at the same sentence months later.

2. STAGE 02MEASURE

   Questions are generated against the concepts found in the text, not against its surface wording. Each item declares which concepts it targets, what dimension it measures, how hard it is, why it was asked, and which blocks support it. An item that cannot answer those questions is not shown.

3. STAGE 03PERSIST

   Your answer becomes append-only learning evidence. Evidence updates mastery for the concepts behind it, and a wrong answer that repeats becomes a tracked misconception rather than a lower score. The assessment is disposable; the graph is not.

Surfaces

WEB APP

Paste text or upload a PDF at /start, work through the items, and read the graph that accumulates across everything — per concept, not per document.

EXTENSION

A thin Chromium adapter that extracts the page or selection in front of you and hands off. It carries no heavy work and stores no learner truth. Not yet published — see Status.

Changelog

Dated by owner acceptance of each stage, newest first. A stage is listed only once its gate closed.

1. STAGE 083 August 2026

   Public landing surface

   The public landing page at /, self-hosted brand typography, and a corrected session boundary: visiting the front door no longer creates an anonymous learner. The product home moved to /start.

2. STAGE 0730 July 2026

   Learner-intelligence refinement

   Consistent loading, empty, error, and recovery states across every product route; mobile and accessibility passes at 1440 and 390; the next-learning-action explanation shown with its evidence rather than as a bare score.

3. STAGE 0629 July 2026

   PDF product adapter

   PDF upload up to 40 pages, parsed through a bounded adapter that validates MIME type, file signature, byte size, page count, and extracted-text size before any work begins. Originals are held in private storage with metadata in PostgreSQL, and deletion removes both.

4. STAGE 0524 July 2026

   Quality, ownership, and demo release

   Engine-specific golden sets with an evaluation runner, learner export and deletion, and the first deployed demo release.

5. STAGE 0423 July 2026

   Identity without friction

   Anonymous sessions that need no signup, and Google linking that transfers ownership of existing work transactionally — a cancelled or failed sign-in cannot delete anonymous work.

6. STAGE 0322 July 2026

   First product surface

   The Chromium extension shell and its handoff contract: single-use tokens, hashed at rest, expiring in ten minutes, with material content kept out of URLs and logs.

7. STAGE 0220 July 2026

   Canonical learning loop

   Capture, measure, and persist end to end — material versioning into ordered source blocks, concept extraction, assessment design, grading, append-only learning evidence, and mastery projection.

8. STAGE 0119 July 2026

   Reproducible foundation

   The pnpm workspace, PostgreSQL 17 with committed Drizzle migrations, shared Zod contracts, typed startup environment validation, and the GET /health readiness check.

Evaluation method

A fluent answer from a model is not evidence of quality. Output is accepted only when it is structured, grounded, measurable, and inside its budget.

What every item must carry

Generation is not one large prompt. Concept extraction, assessment design, and focused evaluation are separate engines, each with named input and output schemas, a versioned prompt identifier, a timeout, a repair policy, explicit failure codes, and its own golden set. No engine returns prose for another engine to parse, and no engine writes another engine's output.

Grading, mastery updates, and the initial next-action choice are deterministic. They do not call a model at all.

The four dimensions

Difficulty is not the only axis. What a question is trying to find out is recorded with the answer, so weakness can be attributed to a kind of thinking rather than to a topic.

RECALL

Retrieve a stated fact, term, or value. Cheap signal, weighted low.

REASONING

Explain why something follows — mechanism, cause, consequence. The highest-value signal, and the main source of misconception candidates.

APPLICATION

Use the idea on a situation absent from the source text. Updates transfer confidence on related concepts.

DEFINITIONS

Draw the boundary — what counts, what does not. Updates concept boundaries and the relationships resting on them.

Grounding

Every item carries the source blocks it came from. That citation is not decoration: it is how you check a question against the text, and how you catch the model being wrong. Golden fixtures deliberately include the hard cases — ambiguity, dense technical prose, converted tables, negation, and multilingual material — and a regression against them blocks the change.

Status

GraspN is pre-launch. This is an honest build state, not an uptime dashboard — there is no incident history to publish yet.

ENVIRONMENT

Staging only, on controlled synthetic data. Managed PostgreSQL is required before public launch or any meaningful learner data.

SHIPPED

The full capture → measure → persist loop, pasted text and PDF up to 40 pages, anonymous sessions with optional Google linking, the knowledge graph, learner export, and account deletion.

NOT YET

The browser extension is built but unpublished, so "Add to Chrome" is inert. There is no billing, no team surface, and no public API.

READINESS

The live check is /health, which reports configuration, build, and database readiness for the running deployment.

Privacy

Learner knowledge is the product, which is exactly why it cannot be treated as a disposable cache or held hostage.

What is stored

Your identity and session; the material you supplied and its source URL; your answers; the concepts, relationships, evidence, misconceptions, and mastery derived from them; and safe operational diagnostics for each generation run. Nothing else about you is collected.

What is not

Provider response bodies, prompts, credentials, and auth tokens are never exported and never shown. Learner material is not used to train a model, and deletion keeps no shadow copy for that purpose. Material text is never placed in a URL or a log.

How access is bounded

Browser, extension, and public route input are all treated as untrusted, and so is model output until it passes schema and quality validation. Authentication establishes who you are, never permission to another learner's graph. Anonymous use is supported without exposing personal information; sessions use secure HTTP-only cookies in production. Extension handoff tokens are single-use, stored only as a hash, expire in ten minutes, and travel in the URL fragment so they never reach the server log.

Data ownership

Your understanding is not our asset. It is readable, portable, and deletable — on request, not on appeal.

EXPORT

One request returns structured JSON containing your materials and source blocks, generation runs with their diagnostics, the canonical concepts and relationships, your misconceptions, append-only evidence, mastery projections, assessment history, and retained PDF metadata. Unknown model pricing exports as null, never as a guess.

DELETE

Deleting your account requires an explicit confirmation and a client-generated token whose hash alone is stored. In one transaction it removes your materials, graph, evidence, mastery, assessments, identity mapping, and auth user, then removes private file originals. It removes the graph, not just the session.

EXPLAIN

Every item can say what it measured and why it was asked, and every next learning action states the evidence behind it. There are no unexplainable scores.

Manage both from your account.

About

Most reading tools help you collect. Very few help you find out whether any of it stuck.

GraspN was built on one distinction: a quiz generator saves quizzes, while a learning layer saves concepts, evidence, misconceptions, and mastery. The extension is the visible surface; the graph underneath it is the actual product, and every engineering decision here — the typed contracts, the append-only evidence, the citation on every item — follows from refusing to blur the two.

Contact

Questions, defects, or data requests: [email protected].