Make a sample backup and prove you can restore it

Run the exact sample in your own disposable folder

Only the location and available checking tools change in this route. Do not substitute personal files, paths inside the archive, content, manifest or ZIP format. Use an existing owned disposable folder with no real data; keep input files, contents and directories stable. Download and reopen the complete supplied files, preserve originals, and start a separate chat with the prompt below. For a code review, send:
“Review the COMPLETE SAVED sample-backup.py below against the exact five-file contract we just confirmed. Trace physical and uncompressed limits, exact names/order, paths/duplicates, regular metadata/flags, declared and actual sizes, manifest, independently known bytes and final canonical archive comparison before any restore output. Keep descriptor-relative/no-follow requirements; no broad extract-all or weakening for another platform. Explain the read-only inspect command and preview commands; do not claim execution. COMPLETE REOPENED UTILITY AND FIXED CONTRACT: [paste full source and full contract]. ACTUAL PLATFORM/INSTALLED PYTHON: [state or unknown].”
Read the source yourself and obtain the required independent implementation review before any write; record what remains pending. If the reviews and compatible installed runtime are available, work from the disposable downloads folder. First inspect with python3 -B sample-backup.py inspect --archive ./sample-backup.zip. Then send:
“Use only MY EXISTING OWNED DISPOSABLE DOWNLOADS FOLDER for this exact synthetic fixture. Give previews for create-sample to an absent sample-originals, optional pack to a different absent sample-created.zip, and restore from the unchanged sample-backup.zip to absent sibling sample-restored. Explicit --write is required for creation; previews and inspect do not create files. Before each write confirm that required reviews are complete and the chosen name is still absent. No scans, real files, overwrite, deletion, cleanup, install, network or upload. Keep all chosen inputs, contents and directories stable; concurrent modification is unsupported. Explain disclosed partial NEW output on an I/O failure and no rollback. ACTUAL DISPOSABLE FOLDER, PLATFORM, REVIEW STATUS AND INSPECT RESULT: [provide]. COMPLETE UTILITY AND FIXED CONTRACT: [provide again].”
Use the exact commands in sample-backup-usage.md; if choosing different absent sibling names, change every later reference consistently. Review each preview before adding --write. The optional pack uses --archive ./sample-created.zip; restore uses the original ./sample-backup.zip. Do not retry into an existing or partially written destination. If execution is unavailable, stop at explanation with writes not run. After an actual successful restore, send:
“Review the COMPLETE SAVED sample-backup-check.py independently from the utility. It must read only named originals, restored files, unchanged archive and separate manifest and compare exact literal bytes, lengths and hashes, without writes/scans/repairs. Supply the command using MY actual directory names. Run only if authorized tools can access those exact local files; otherwise I will run it locally and paste its complete output. Do not infer access from filenames. COMPLETE CHECKER: [paste]. ACTUAL ARCHIVE/MANIFEST/ORIGINAL/RESTORED LOCATIONS AND TOOLS: [state]. ACTUAL RESTORE OUTPUT OR FAILURE: [paste].”
If permitted, run python3 -B sample-backup-check.py --archive ./sample-backup.zip --manifest ./sample-backup-manifest.sha256 --originals ./sample-originals --restored ./sample-restored with your consistent names. Save its actual output to a new report and reopen it. On a mismatch keep all files unchanged. Then send:
“Review the COMPLETE BUNDLED SYNTHETIC TEST SUITE and utility below. Use only an explicitly chosen existing disposable scratch folder; keep the downloads and actual originals unchanged. Run python3 -B sample-backup-tests.py --scratch-parent [my existing scratch folder] only with supported authorized execution. It creates and retains new isolated synthetic copies. Check invalid-input refusals create no restore output, and separately distinguish the injected mid-write error’s retained partial NEW output. No deletion, cleanup, overwrite, scans of user folders, install or real files. Record actual environment, date, code/archive hashes, executed cases and failures; leave unrun checks pending. COMPLETE UTILITY, TEST SUITE, FIXTURE AND SAVED BYTE-REPORT: [provide]. EXISTING DISPOSABLE SCRATCH FOLDER AND AVAILABLE TOOLS: [state].”
These are the exact sample’s checks, not a route for real-data backup. Keep the results and any partial new output; do not silently recreate originals. Return to the sample chat for the original advanced audit with the complete code, fixed fixture, hashes and actual test outputs; no adversarial-concurrency, crash-atomicity or production protection is established.

Explain how to run the EXACT supplied synthetic backup/restore exercise in my own existing owned disposable folder. Do not use my real files or generate a different fixture. First confirm this full fixed payload: notes/readme.txt is hex 53616d706c6520617263686976652e0a (16 bytes); notes/café.txt is hex 4f6e6520636166c3a9206e6f74652e0a (16 bytes, NFC filename); data/counts.csv is hex 6974656d2c636f756e740a63617264732c31320a70656e63696c732c340a (30 bytes); empty.txt is zero bytes; marks/pattern.bin is hex 00017fff (4 bytes). Text is UTF-8/LF. Exactly five payloads/66 bytes, plus the exact separate 407-byte manifest inside the canonical 1,125-byte stored ZIP. Archive SHA-256 is 0fbf539d4ebb19c1042fdd3486d297a6b151572afd3b516722943a3c3b6f3b0b; manifest SHA-256 is 28fda419143b9e271247e3a63083bacbcc61f138859c427964fac9f032d5496a. Physical bound 96 KiB; total uncompressed bound 64 KiB including manifest. Accept only the exact supplied canonical encoding, not arbitrary ZIPs within these bounds. Validate all names/metadata/bytes before output creation. Preserve downloaded archive and originals; use exclusive new files and an absent sibling restore directory. No scans, real files, overwrite, deletion, cleanup, network or install. Inputs, contents and directories must remain stable; concurrent modification is unsupported. A write failure may leave disclosed partial NEW output; no rollback. Require user code review and independent implementation review before writes. Ask about missing tools/review/folder information; begin with explanation only and do not claim a restore has occurred.

MY EXECUTION CONTEXT
[State the existing owned disposable folder containing the complete downloads, that it contains no real data, the actual platform and already-installed Python version or unknown, and code/independent-review status. Paste the complete sample-backup-fixture.json and fixed contract for checking. A path alone does not let an AI read local files. If no supported runtime or review exists, keep execution pending.]

Check your own result

• Only your disposable output location changes. The five paths, exact bytes, canonical archive and manifest remain the supplied fixed fixture; different real data needs a separate design and permission.
• User and independent code review precede writes; existing destinations and unsupported platforms are refused. Every new location stays within the intended disposable area, with originals/restore as siblings.
• Actual restored files are read and compared with independent literal byte controls. A listing, preview or historical result cannot substitute; failures never trigger repairs to originals.
• Inputs/contents/directories remain stable. Invalid-input refusal and mid-write failure are distinct; partial NEW output is disclosed and retained with no deletion or atomic-rollback claim.
• Only performed checks have results. No real-data recovery, hostile-concurrency guarantee, retention, encryption, off-device protection, platform-wide safety or consumer-AI performance is established.