Data layer at tag v0.3.0-pre.4: citable snapshot named in CITATION.cff, verification paragraph states a tag-to-tag comparison

CITATION.cff

@@ -15,7 +15,7 @@ authors:
     # affiliation is to be added here or in any listing, archive record or citation form
     # Do not attempt to resolve the handle to a person.
 type: dataset
-version: "0.3.0-pre.2"    # matches the tag v0.3.0-pre.2; pre-release, definitions may still move
+version: "0.3.0-pre.4"    # matches the tag v0.3.0-pre.4, the newest citable snapshot; scripts/preflight.sh --release <tag> refuses if they differ
 date-released: "2026-09-11"   # last revision of the document set, not a release date
 # doi: ""                 # dataset DOI, then paper-page DOI
 repository-code: "https://github.com/2makeitwork/paper_ai_Qos"
@@ -24,7 +24,7 @@ license: >-
   CC-BY-4.0 (text, tables, figures and evidence data; see LICENSE-content.md);
   MIT (executable tooling under scripts/; see LICENSE-code.md)
 # A commit hash cannot be written into the commit it names, so the snapshot to cite is the
-# tag: v0.3.0-pre.2. Zenodo's GitHub integration mints the version DOI from the release page.
+# tag: v0.3.0-pre.4. Zenodo's GitHub integration mints the version DOI from the release page.
 keywords:
   - AI service quality
   - large language model service measurement
@@ -43,6 +43,13 @@ preferred-citation:
   # url: "https://huggingface.co/papers/..."   # paper page once published
   institution: "Independent"
 versions:
+  - date-released: "2026-09-11"
+    version: "0.3.0-pre.4"
+    changes: "Pre-release line. Methodology v0.1 with Case Study 1: paper section 1.4 rebuilt as a
+      category table, the model-authenticity works recorded as considered-and-not-cited, the
+      capability-versus-delivery asymmetry stated in section 1.1, an unaffiliated byline, dated
+      pre-release notices, and a release gate (scripts/preflight.sh) that checks the published
+      artifact rather than the author's working copy"
   - date-released: "2026-09-10"
     version: "0.2.0"
     changes: "Round 2 collection: evidence/incident_20260910/ (62 request episodes, per-request context occupancy, the compaction pair and three negative results)."

README.md

@@ -78,24 +78,24 @@ recomputed from these files rather than taken on trust.
 
 ### Verification after publication
 
-The whole published repository was downloaded back to an empty directory and compared file by
-file with the working tree: **34 of its 36 files were byte-identical** (measured 2026-09-11; the
-repository's commit history shows which revision that was). The two exceptions are
-known and intended — the hub's own `.gitattributes`, created with the repository, and
-`README.md`, which is this card with three lines of internal instruction removed (the source
-repository's `README.md` is a different document, its homepage). The generated table ends with
-a line carrying a content hash and a second-precision timestamp: the hash is what a machine
-compares, the time is what a person reads, and when the statistics are unchanged the script
-leaves the file alone — so running the checks cannot dirty a checkout or produce new bytes, and
-two copies with the same hash hold the same statistics whatever their stamps say.
-(Incidental finding for anyone scripting the client: `hf download --include` accepts repeated
-patterns but only honoured the last one, so pass one pattern per invocation.)
+The whole published repository was downloaded back into an empty directory and compared file by
+file with the tag it was built from — which is exactly what `scripts/check_release_sync.py` does,
+so this is re-checkable by a reader rather than autobiographical. Measured 2026-09-11 against tag
+`v0.3.0-pre.4`: **every data-layer file agrees byte for byte**, `README.md` included — that file
+is this card with its three internal-instruction lines removed, and the source repository's own
+`README.md` is a different document, its homepage. The hub holds one file no tag contains: the
+`.gitattributes` it created with the repository. The generated table ends with a line carrying a
+content hash and a second-precision timestamp: the hash is what a machine compares, the time is
+what a person reads, and when the statistics are unchanged the script leaves the file alone — so
+running the checks cannot dirty a checkout or produce new bytes, and two copies with the same hash
+hold the same statistics whatever their stamps say. (Incidental finding for anyone scripting the
+client: `hf download --include` accepts repeated patterns but only honoured the last one, so pass
+one pattern per invocation.)
 
 Both scripts were then executed **inside that download**, with nothing else present:
-`scripts/verify_dataset_card.py` exits 0 and `scripts/analyze.py` prints a PASS line per check
-and fails none. Its count is lower there than in the source repository, because the checks that
-compare the paper with the case report have nothing to read: they print an explicit skip, never
-a silent pass.
+`scripts/verify_dataset_card.py` exits 0, and `scripts/analyze.py` fails nothing — printing a skip
+for the checks that compare the paper with the case report, since those narratives live in the
+source repository and their absence is announced rather than passed over as a pass.
 
 ### Composition