Scoped save-diagnostics package evidence
The release validation and run instructions
define the result and its limits. Desktop source is
0e83157e09d52e6152df91208e41693b62da2d46, packaged documentation
ef28f1784efe6717b0ca72c0eb3c97bc30ebfe47, and the reused server source is
715036e132677dd2bc2fb37ed316d93dc9987aec. This is a scoped Windows development
preview, not complete OA/SA acceptance or the separately owned follow-up union.
The exact ZIP is 34,191,561 bytes, SHA256
8b4e353e075f22350a07794ee47a02cf85725501223299cf367134aa1f5b63b7.
The package itself contains the binaries, notices and documentation; this packet
contains the portable verification evidence, not duplicate executable archives.
Confirmed publication receipts identify release 9 and all three
downloaded assets; they are outside the preceding raw inventory.
Navigation
- Independent source/artifact/native/REST/claims review.
- Independent packet/index closure review, including exact reconciliation of all 162 copied artifacts with their original files.
- Artifact precheck by the preparation-helper author; this is separate from the independent final review above.
- Parent native observation, 19 captures and executed commands.
- Complete source/model comparison, fresh server assertions, and valid-model warning reconciliation.
- Extraction receipt, build provenance, and Git-identified source snapshot.
- Collection identities and byte inventory: 164 indexed files, 15,568,173 bytes, plus the inventory itself. Raw scoped attributes preserve captured bytes.
Reproduce the offline verification
From this repository checkout, with Python 3:
python harness/experiments/save-diagnostics-package-20261001/raw/verify.py
python harness/experiments/save-diagnostics-package-20261001/raw/verify.py --current-source
The first checks exact packet membership/bytes, pinned Git source identities,
package/extraction input identities, native/process receipts and complete
37-file/30-source-file/full-Project comparisons. The second additionally requires
the current source files to match the packaged snapshot; it may correctly fail
on a later implementation. After staging this packet, add --index to verify
Git's staged blob bytes exactly, including ignored logs explicitly staged for
this evidence. These checks do not run the archived native/REST/build scripts.
The parent's offline and staged-index checks passed with all 637 source inputs
unchanged. Initial diff checks flagged captured CRLF bytes, followed by intentional
terminal/HTTP blank lines. Scoped raw-only whitespace attributes preserve those
original bytes; production/doc formatting policy and the inventory are unchanged.
Historical scripts retain their original Windows paths and exclusive output directories. Use the linked release instructions and explicit fresh-directory adaptation for runtime reproduction. Never replay the injected-comment helper against a user's model. Stage/archive integrity alone is not native acceptance.
Preserved boundaries and failed assumptions
The original server build used LLVM-MinGW 20260908; the current package selects the 20260826 runtime. Recovered public donor records retain their original identities, and the fresh extracted smoke validates the exact current pair. Missing historical local paths are preserved as missing; portable recovered copies supply the evidence. The old helper-review's missing-provenance concern is superseded by that recovery, not erased.
Review-script assumptions about null registry paths, JSON wrapper depth, CRLF count and Windows text decoding were corrected and retained in their reviews. The initial parent oracle-adaptation replacement-count failure occurred before output creation; only its occurrence count was corrected. The premature native loading capture remains, and is not used as loaded-model proof. The corrected prior Unicode02 smoke helper is the byte-identical donor; its separate failed predecessor is not claimed as passing.
Seven tutorial companions were copied by the parent; Save As wrote 30 source files. The complete-model oracle uses the same core implementation family, so it verifies structural conservation rather than independent language semantics. Forced server exit 1 is intentional, not graceful/power-loss proof. Two desktop exit-0 sessions on one account/machine with restricted PATH do not establish clean-machine, keyboard accessibility, minimum-window, broad DPI, Linux or whole-release acceptance. The native error footer truncates; pointer tooltip access was exercised, keyboard-only access was not. A clean model with an external disk conflict still says All changes saved in its local-edit header; the explicit failure footer provides the save result. No automatic reload or overwrite is inferred. Existing published releases remain frozen.