InkDraft

SYNTHETIC SOFTWARE DEMONSTRATION ยท NOVEMBER 18, 2026

Read. Edit. Request another read.

A page edit and a fresh read are two separate events. This small software test shows the difference using five generated pages and InkDraft's document-reading tools.

Synthetic test, not an assistant conversation. No physical iPad, handwritten content, consumer AI model, cloud sync, or account connection was used. The steps below are a diagram of a local in-memory test, not screenshots or quoted AI answers.

One document, two reads

  1. FIRST READ

    Request all five pages.

    The test client called get_full_document and followed pagination. Two batches returned all five original generated PNGs; the test compared their byte hashes.

    12345
  2. EDIT

    Replace the image on page five.

    The test changed one generated image and committed a new document revision. An attachment or transcription from the first read would still describe the earlier content.

    12345
    changed
  3. SECOND READ

    Request changes from the first revision.

    The client called get_document_changes. It received page five's updated image, a newer revision, and none of the four unchanged images. The returned image hash matched the replacement, not the first read.

    5
    new image

What was actually checked?

  • All five images arrived across two first-read batches.
  • The second read returned exactly one changed page with the new image bytes.
  • A further read with no change returned no images.
  • A request for the old revision was rejected instead of mixing versions.
  • Switching documents required a new full read.
  • Read operations did not mutate stored document content; editable native ink was not returned.

The recorded run passed all eight assertions on November 18, 2026 at 00:40 UTC. It used the MCP SDK's in-memory transport, InkDraft's document store, and its document tool registration. The four document tools were tested; the production account-profile tool and authorization were outside this exercise.

Read the original synthetic test receipt (JSON text). The receipt contains test checks and limitations, with no customer content or account credentials.

What this does not establish

This checks a software path for fresh page delivery. It does not prove that ChatGPT, Claude, Gemini, or another model will choose the right tool, inspect an image correctly, or give sound feedback. It does not measure iPad sync latency, setup success, accuracy, learning outcomes, or time saved. There were no real-user sessions or acquisitions in this test.

A fresh screenshot could also provide the corrected page. No upload-versus-connection benchmark was run. See when a screenshot is enough.

What to check in your own workflow

With a compatible, separately authorized assistant, first ask it to identify the current document and read a small, non-sensitive detail. Change that detail on iPad, wait for Synced, then ask for another read without putting the new answer in the prompt. Compare the response to the page and, where available, the tool's revision and source evidence.

That is a suggested personal check, not a result we have observed. A matching answer alone does not establish a fresh fetch if the answer was already in the conversation. Avoid private notes while checking a new connection.

To independently examine the workflow, use the original synthetic reproducibility pack: blank PDFs, generated reference strokes, expected observations, and an empty report. Its six product cases remain not_run until a reviewer records an actual authorized test.