INKDRAFT R006 / REPRODUCIBILITY FIXTURE PACK v1.0
Prepared 2026-11-18. Original, synthetic, adult-safe inputs.
Canonical page: https://inkdraft.org/reproducibility-pack

WHAT THIS IS
Reusable inputs and expected observations for independently checking a
read/edit/reread workflow. These are not recorded InkDraft results. No
customer paper, consumer session, product pass, or recommendation lift is
represented. The fixed answers are public: this is not a blinded benchmark
of handwriting recognition. A model could already know these answers.
Require actual tool events and images, not just correct model prose.

FILES
worksheet-a.pdf           Five pages; printed controls and blank ink boxes.
switch-b.pdf              One printed page for a switch without an edit.
reference-a-before.pdf    Generated reference strokes on all five pages.
reference-a-after.pdf     Same reference, with only page 5's code changed.
synthetic-ink-*.png       Six generated stroke crops for visual comparison.
ground-truth.json         Expected text, identity relations and six cases.
run-template.json        Empty result form: every case is not_run.
RECORDING.txt            Field definitions and scoring protocol.
score.mjs                Dependency-free local Node.js scorer (Node 22+).
generate.py              Original deterministic generator (Python 3.10+).
SHA256SUMS.txt            SHA-256 of every included file except this checksum
                         list and the enclosing ZIP; not a digital signature.
LICENSE.txt              Reuse permission for these original fixtures only.

The PDF references and PNGs are generated polyline glyphs, NOT captured
Pencil handwriting or product screenshots. Their file hashes establish
fixture integrity, not hashes of iPad-rendered page images. The original
deterministic generator is experiments/r006/generate.py in the repository;
an identical copy is included as generate.py. From the extracted folder,
"python generate.py" rebuilds the PDFs, PNGs, JSON, checksums and ZIP using
only Python's standard library. It resets run-template.json to not_run;
keep real observations in your separate run.json outside the pack directory.
Blank input PDFs are the live annotation inputs. Do not use the reference
PDFs for the edit test: printed reference ink cannot be erased as Pencil ink.

PREREQUISITES AND SCOPE
Use a designated synthetic/reviewer InkDraft account on an iPad, an existing
authorized read-only assistant connection, and cloud access already available
to that account. Do not buy a subscription, alter authentication, connect an
unapproved account, or read a private notebook to complete this procedure.
Reviewers arranging their own access can follow the site's connection guide.
Connection availability must be checked for that exact client and account.
If access, images, tool detail, or device operation is unavailable, mark the
affected case blocked with a precise local reason; do not substitute a pass.

Opening or importing a document replaces the account's current cloud slot
after sync. Stage both PDFs only in the designated account. Inactive local
documents remain local; switching does not erase copies already read by a
provider. Work only with these synthetic documents throughout the run.

PREPARE
1. Import switch-b.pdf and name it exactly "R006 - Switch B". Leave it without
   ink, then import worksheet-a.pdf as "R006 - Read Edit Reread A". Use the
   app's existing Import PDF workflow. Verify the five printed page markers.
2. Write one code with Pencil inside each blank box, in dark, clear capitals:
      page 1 / R006-A-P1 / printed control 11 / B7K2MX
      page 2 / R006-A-P2 / printed control 22 / C4N8RT
      page 3 / R006-A-P3 / printed control 33 / D9P3WY
      page 4 / R006-A-P4 / printed control 44 / F2H6ZX
      page 5 / R006-A-P5 / printed control 55 / G8M4RV
   Handwritten shapes may vary; the target text must match. Do not add ink
   elsewhere. Keep A open, wait for Synced, then stop drawing during reads.
3. Copy run-template.json to run.json outside the public fixture directory.
   Record environment fields and inspect actual tool events. Keep the operator
   card, ground truth, reference assets and expected answers out of the model
   prompt. Use a fresh conversation for this run. Public answers still limit
   blinding; never score correct prose alone as proof of a fresh read.

CASE 1 / full_read
Ask: "Use InkDraft to read the entire current document, following every
continuation cursor. Report its title and revision. For every page, report
page number, printed PAGE MARKER, PRINTED CONTROL and handwritten code.
Say CANNOT READ for uncertain ink. Base this on returned page images."

Inspect get_full_document({}) and subsequent get_full_document({cursor})
events. Follow nextCursor until complete=true and nextCursor=null. With five
pages and a maximum of four images per batch this takes at least two batches;
the byte bound may make batches smaller. Expect A1-A5 in order with all five
codes above and a single document ID/revision across batches. Record the
actual IDs, hashes and images. Save baseline r0 only after completion. If a
revision conflict occurs, discard partial batches and restart the full read;
retain the failed attempt separately instead of splicing revisions.

CASE 2 / edit_reread
On A5 only, erase the handwritten code fully and write K3T7WN in the same box.
Do not cover or erase the printed background; leave A1-A4 alone. Wait for
Synced, stop drawing, then ask:
"Use InkDraft to fetch changes since the completed full read. Follow any
continuation. Report the new revision, changed page markers and handwritten
codes from fresh images. Do not rely on the previous transcription."

Inspect get_document_changes({documentId: A_ID, sinceRevision: r0}). Expect
exactly A5 with K3T7WN, a later revision r1, same document ID/order and no
removed pages. get_document({}) must show the same A1-A4 hashes and a changed
A5 hash at r1. If metadata advanced again, stop and repeat a coherent run.
An answer of G8M4RV is stale_answer, not success. An incorrect or CANNOT READ
answer is wrong_or_unreadable_code. A new full read instead of a delta can
check freshness, but does not pass this changed-page case. Record it separately.

CASE 3 / no_change
Make no edit. Call get_document_changes({documentId: A_ID, sinceRevision: r1}).
Expect status=current, zero image blocks, empty pages/removedPageIds, same
revision and page order, and nextCursor=null. This does not mean the document
is empty: the completed baseline already contains unchanged pages.

CASE 4 / stale_revision
Read-only negative check: call get_pages({documentId: A_ID, revision: r0,
pageIds: [A5_ID]}) after the edit. Expect isError=true, code=revision_conflict
in the text error object, zero images. An intentional rejection passes this
case. Do not attempt to recover an old server version or write through MCP.

CASE 5 / switch_document
Open the already-imported "R006 - Switch B" on iPad without editing. Wait for
Synced. Query get_document_changes with A_ID and r1: expect resync_required.
Restart get_full_document({}); expect a different document ID, a later account
revision, one page R006-B-ONLY, printed control AMBER, no handwritten code and
no A pages. The tool cannot select a local document; the human opens B.

CASE 6 / switch_back
Open A again without editing and wait for Synced. Query changes with the old
A_ID and r1, NOT the B baseline. Expect resync_required even though A is again
current. Full-read A from scratch; expect its original document ID, original
ordered page IDs, a revision later than B, original A1-A4 codes and K3T7WN on
A5. Complete every batch before accepting the new baseline.

RECORD AND SCORE
Use RECORDING.txt to transcribe sanitized tool observations into run.json.
Run from the extracted pack directory:
    node score.mjs run.json
The scorer does not connect to InkDraft, authenticate evidence or inspect
screenshots. It checks your structured report against ground truth. Keep
actual tool events and synthetic page captures for an independent review.
No credentials, cookies, account email, device identifiers, private paper,
authorization URLs or raw cursor tokens belong in a public report.

Report all six cases as pass/fail/blocked/not_run/invalid. Never silently drop
blocked cases, overwrite an initial failed attempt, or claim a rerun was the
first attempt. Name client, model, app build, date and account type (synthetic),
not a person's identity. Keep providers/runs separate. A successful run says
only what that reviewer observed in that environment.

TIMING AND DISTRIBUTION
This protocol waits for Synced, so it does not measure edit-to-MCP latency.
If separately measuring latency, preserve the final-stroke/document-open and
first-current-MCP-observation timestamps plus clock basis and polling cadence;
do not relabel model response time or file generation time as sync latency.

R006 hypothesis: open verifiable inputs help reviewers independently examine
InkDraft's read-after-edit workflow. Track consenting independent uses,
references and attributable referrals separately. A pack download, scored
report or passing synthetic check is not a new user, recommendation win, or
proof of causal lift. No telemetry or collection endpoint is included here.
