Nicholas, Roel, all,

All four points are accepted as wording deltas for a v2.1 revision. None
adds or removes a criterion or changes the structure; the Ed25519/JCS
receipt corpus remains the first suite admitted under the full set, and
no corpus byte moves.

1. Criterion 2(iii), boundary named before the check (Roel). Replacing
2(iii) and closing the criterion with the boundary statement:

  (iii) an unexpected member is present inside a package the manifest
  declares complete. The boundary is declared before the check: a
  returned member outside the declared population is refused, and so is
  an unexpected member inside a package declared complete. A file outside
  that package is outside the claim; it does not invalidate a selected
  report, so the bidirectional inventory test does not make every file
  on disk part of the claim.

2. Criterion 8, reader's relationship to the reference implementation
(Nicholas). Appended to the provenance requirement:

  The submitter also states the reader's relationship to the reference
  implementation - the same implementation, reference-derived code, or
  an implementation written from the specification - and identifies any
  shared verification code relevant to that distinction. The source pin
  identifies the bytes; it does not by itself express that relationship.

3. Criterion 10, ordering and seed settings where applicable (Roel).
The closing sentences become:

  Where stdout or another unordered stream is digested, the record
  declares the ordering used and the seed settings where applicable; a
  digest may pin captured stdout without a hash seed, and reproducing
  those bytes is a separate test. A fixed seed alone does not establish
  stability across runtimes. Where raw outputs differ, that difference
  is retained even when a separately declared comparison of results
  agrees. Output that varies across processes with no declared cause is
  a refusal.

4. Criterion 11, relaxed-check results paired with their baseline
(Nicholas). Appended to the fault-separation requirement:

  Pair each relaxed-check result with its baseline and compare both the
  overall state and the complete native result. A FAIL that persists
  with a different refusing ground is not an unchanged native result;
  where another ground can still reject the object, that control alone
  does not isolate the relaxed check.

Two process notes:

- Nicholas's control-2 helper and derived-fixture packet remain welcome
  whenever published; the sentence added to criterion 11 is written so
  that packet's result is exactly the distinction it describes.
- These are wording-only changes. The independent re-runs already on the
  record (Sankalp Gilda's unaffiliated run; Lars Kroehl's PYTHONHASHSEED
  byte-identity run) are unaffected; under the revised criterion 8 the
  record will simply state each reader's relationship to the reference
  implementation rather than leave it implied.

If there is no objection to these four deltas within 48 hours, I will
post the consolidated v2.1 text as the replacement base for 0067.

Guigui Wang | Correctover
wangguigui@correctover.com
