R006 v1 / STRUCTURED OBSERVATION FORMAT

Use run-template.json; the scorer accepts only schemaVersion 1, packId
inkdraft-r006-v1, evidenceType reviewer_report. Node 22+; no packages needed.
Keep a separate run.json per attempt/client. Do not publish the template as a
result. All codes are strings; empty/absent ink on B is JSON null.

run fields:
dateUTC: YYYY-MM-DD of the actual run; client/model/appBuild/iPadOS: observed
versions (use "not exposed by client" for a model version genuinely hidden).
syntheticAccountConfirmed and toolEvidenceInspected must both be true for an
observed case to be scored. Correct prose without tool/image evidence is not
enough. Do not set either flag merely because this document says to do so.

Each of the six cases has status not_run, blocked or observed. A blocked case
requires a precise reason. The scorer assigns pass/fail; never enter pass.
The scorer suppresses freeform reasons from its output; retain them in your
local report for review. Missing cases remain not_run in the denominator.

Use actual document/page IDs or a consistent, reviewer-verifiable pseudonym
mapping. Preserve account revision numbers and page hashes as returned by
InkDraft. Hashes are lowercase 64-character hex. They are payload hashes,
not expected hashes of these PDFs or reference PNGs. Never invent them.
Replace nextCursor tokens with nextCursorPresent: true/false. Do not record
raw cursors or authentication headers. imageCount counts actual image blocks.
marker/printedControl/code are the assistant's transcriptions checked against
the returned images. If uncertain, use "CANNOT READ". Do not fill in answers
from the operator card when the assistant failed to read them.

CASE full_read: observed has a batches array. Each batch object contains:
  status, documentId, revision, title, pageCount,
  complete, nextCursorPresent, imageCount, pages
Each pages entry contains:
  id, hash, pageNumber, marker, printedControl, code
All fields except marker/printedControl/code are transcribed from tool
results (nextCursorPresent derives from whether nextCursor is non-null).
For the handwritten code only, scoring ignores whitespace and letter case;
it does not normalize punctuation or confusable characters. Printed markers
and controls match exactly. A partial or mixed-revision read fails.

CASE edit_reread: observed contains:
  sinceRevision (r0 used in the request), status, documentId, revision (r1),
  title, pageOrder (all 5 IDs), removedPageIds (empty array),
  nextCursorPresent (false), imageCount (1),
  pages (one A5 entry: id, hash, marker, printedControl, code),
  currentPageHashes (all 5 hashes, in order, from get_document at r1).
This one-page delta needs one batch. If extra edits produce other changed
pages, record the failure; do not hide them to match the expected count.

CASE no_change: observed contains the same delta fields as edit_reread,
except sinceRevision is r1, revision stays r1, imageCount is 0, pages is [],
and currentPageHashes is not needed.

CASE stale_revision: observed contains:
  requestDocumentId (A), requestRevision (r0), requestPageIds ([A5_ID]),
  isError (true), code (revision_conflict), imageCount (0).
The code comes from the JSON text content of the error response. This case
requires visible request arguments and error details. Otherwise mark blocked.

CASE switch_document: observed contains:
  openedWithoutEdit (human observation), requestDocumentId (A),
  sinceRevision (r1), oldBaselineStatus (resync_required),
  full: {batches: [...]} in the full_read format, this time for B.
The uninked B page has code: null. Its full read has one page/one batch.

CASE switch_back: same fields as switch_document, still requestDocumentId A
and sinceRevision r1 (the OLD A baseline), but full now contains all 5 A pages.
Full read uses new revision numbers from actual events and original A IDs.
After A-B-A the page hashes need not be used as evidence of transport reuse.

The cases depend on a completed coherent first read and edited baseline.
If an earlier case cannot establish its baseline, mark dependent cases
blocked. An observed dependent case without a passing prerequisite fails.
Use a new report for any restarted sequence; preserve the earlier attempt.

CLI exit codes:
  0 = all six reported cases pass
  1 = at least one case fails
  2 = invalid input or invalid observations (takes precedence over failure)
  3 = incomplete/blocked run without failures or invalid observations
Output always states that it scores operator reports, not authenticated
product evidence. No pass percentage or recommendation score is produced.

Suggested public report: pack version, date, client/model/app/iPadOS, six
case statuses, precise sanitized blockers, link to consented synthetic tool
evidence, and whether a separate reviewer verified it. Label author-run and
independent runs separately. A repository unit test of score.mjs is synthetic
software validation and never belongs in this product-run template.
