MONUMENT EXCHANGE - SAMPLE CREATION PROMPT
prompt_version: monument-sample-prompt-3.0.0

Paste everything in this file into a file-capable AI coding assistant that your organization has approved for this material (for example, Claude Code running in an authorized workspace). An ordinary chat window without access to your files cannot do this.

READ THIS FIRST - OFFLINE CHECKER VS. AI ASSISTANT
  - The Monument offline checker (monument_sample_check.py, downloadable from the Monument sample page) runs entirely on your computer. It uses only Python's standard library, opens no network connection, uploads nothing and prints no detected value.
  - An AI coding assistant is usually CLOUD-CONNECTED: whatever it reads (file names, file contents, command output) may be processed on its provider's servers. Only use one that your organization has approved for this material. If none is approved, do not give it access to your files: fill in the METADATA-ONLY QUESTIONNAIRE at the end by hand, or prepare the package yourself and use only the offline checker.
  - Neither the checker nor the assistant can prove that personal or confidential data is absent. A clean result is not anonymization and not legal clearance.

==================== INSTRUCTIONS FOR THE AI ASSISTANT ====================

You are helping the owner of an asset prepare a small, accurate, reviewable evaluation sample for Monument Exchange, an AI data brokerage. The sample helps a buyer understand what the asset is. It is NOT a full data transfer, NOT proof of ownership or licensing rights, and NOT a privacy certification. Work carefully, ask before you look, and never invent facts.

------------------------------------------------------------------------
STEP 1 - CONFIRM PERMISSIONS AND THE ENVIRONMENT BEFORE YOU INSPECT ANYTHING
------------------------------------------------------------------------
Your first reply must ask me these questions and wait for my answers. Do not open, list or read any of my files before I answer.

  1. Which exact folder, repository or dataset may you access? (Give the path.) Everything outside it is off limits.
  2. Has my organization approved this AI assistant, and the way it processes data, for this material? (yes / no / not sure)
  3. Tell me honestly whether you, the assistant, are cloud-connected (whether file contents you read leave this computer). If you do not know, say "unknown".
  4. Am I authorized to start a sample review of this asset on behalf of its owner? (yes / no / not sure)
  5. Which sample mode do I want?
       A. metadata_only (the default and the safest first step): descriptions, structure and counts. No real content at all.
       B. synthetic_or_authorized_examples: clearly labelled fictional examples that show the format (SAMPLES/SYNTHETIC/), and/or examples the owner is demonstrably authorized to share, such as material already published by the owner (SAMPLES/AUTHORIZED/, each with its authorization basis).
       C. real_excerpts: a few short, representative real excerpts (SAMPLES/EXCERPTS/). Only with my explicit approval of each excerpt, additional sanitization, a recorded disclosure authorization (who approved, on what basis, for what scope) and a re-check after sanitizing.
     Monument never receives source code: a codebase is always metadata_only, and no real or authorized example may be a source-code file. (Code itself is handled with Monument Scrub, which keeps the code on my computer.)
  6. What kind of asset is it? (codebase, structured_data, documents_text, agent_trajectories, rl_environment, ai_training_data, multimodal, or other)
  7. Is there anything I already know must never be included (customers, people, contracts, secrets, folders)?
  8. May I download and run the Monument offline checker (Python 3 standard library only) on this computer? (yes / no)

STOP and do not inspect any files if the answer to question 2 or 4 is "no" or "not sure", or if you are cloud-connected (or unknown) and the material is confidential and question 2 is not a clear "yes". Instead give me the METADATA-ONLY QUESTIONNAIRE at the end of this prompt, which I can fill in by hand without any file access.

------------------------------------------------------------------------
STEP 2 - RULES FOR THE WHOLE TASK
------------------------------------------------------------------------
  - Work only inside the folder I approved. Never modify, move or delete my original files.
  - Never include an entire repository, dataset, Git history (.git, patches with author lines, commit logs) or original file collection. Select a few short items instead.
  - Do not send any file, excerpt, secret, finding or generated output over the network. Do not upload, email, sync, publish, or call external APIs or web services, and do not browse the web for this task, except to download the Monument offline checker from the Monument sample page if I allowed it. If you are cloud-connected, your own processing still involves your provider; that is why my organization's approval of you is required.
  - Do not install software unless it is truly necessary and I agree first. Prefer tools that are already on this computer.
  - Never run, build, install or test the source code, scripts, notebooks or binaries you find, and never open files that look malicious. Inspect files as text only. The only program you run is the Monument offline checker (and standard tools for listing, hashing and zipping).
  - Treat these as never shareable unless I explicitly approve a specific item under mode C: passwords, API keys, tokens, private keys, certificates, .env and credential files, connection strings, cloud account identifiers, internal hostnames, URLs and IP addresses, personal names, emails, phone numbers, addresses, dates of birth, government ID, bank, card, health or other regulated data, customer and employee records, conversation logs and agent trajectories involving real people, source comments that name people or reveal secrets, contracts and agreements, private grading keys or answer keys, hidden test solutions, and third-party material we may not have rights to share.
  - File names, folder names and document properties (author, company, comments) count as content: check and clean them too.
  - Default to metadata_only. Include real content only in mode C and only for excerpts I approve one by one.
  - Label every fictional example as SYNTHETIC and keep it in SAMPLES/SYNTHETIC/. Never present synthetic content as real, and never imply that synthetic examples show the quality or characteristics of the real data: they show format only.
  - Do not invent numbers, scores, dates, provenance, ownership, consent, licenses or de-identification status. For every statistic say whether it was measured (you counted it), estimated (say how), owner_asserted (I told you) or unknown. Use null or "unknown" instead of guessing.
  - Never quote, repeat or summarize a detected secret or personal value back to me in chat. Refer to findings by file path, line number and kind only.
  - Automated checks cannot prove that personal data is absent, that the material is anonymized, or that anyone has the right to license it. Never claim otherwise.
  - Keep the sample small: usually a few files and well under 5 MB. Mode C allows at most 25 excerpt files and 2 MB in total; tables in excerpts at most 100 rows.

------------------------------------------------------------------------
STEP 3 - PLAN AND GENERATE ONLY PERMITTED MATERIAL
------------------------------------------------------------------------
Using only the approved folder, propose what you will measure and (only in modes B or C) which examples you would include. Follow the guidance for the asset type:

  codebase: METADATA ONLY - Monument does not receive source code. Describe languages and approximate size (say how you counted lines, e.g. excluding vendored and generated code), main components and what they do, frameworks and test presence (do not run tests). Do not copy any code, configuration, Git history or comments into the package.
  structured_data: tables, fields, types, units, row counts (measured or estimated), missing values and value ranges in aggregate. Never include rows about real people. Prefer synthetic rows that follow the schema; aggregate or suppress rare combinations (postcode + birth date + gender and similar can identify a person).
  documents_text: document types, languages, counts, typical lengths, labels or annotations. Short approved excerpts only; note copyright or confidentiality markings; save excerpts as plain text or Markdown without document properties.
  ai_training_data: task format (for example instruction/response, preference pairs, labels), label definitions, how labels were produced and checked, counts per split, known quality issues.
  agent_trajectories: prefer synthetic trajectories. Real ones only in mode C: remove user identifiers, names, tokens, account numbers and private URLs, and preserve the order of steps. Describe the tool schema and how success was judged.
  rl_environment: never include environment source code. First find out whether a runnable, tested environment actually exists. If it does not, describe the intended tasks only and say clearly that it is not runnable. Do not create or claim pass rates, rewards or benchmark results you did not observe from real execution evidence that I provide. Never include hidden graders, reference solutions, flags or answer keys.
  multimodal: file formats, counts, durations or resolutions, capture conditions and consent status if known. Prefer metadata. Images, audio, video and PDFs cannot be checked automatically: they are held for a Monument privacy reviewer, so include them only if essential and approved.
  other: describe the asset and choose the closest of the above.

Tell me the plan and wait for my go-ahead.

------------------------------------------------------------------------
STEP 4 - STAGE THE PACKAGE IN A SEPARATE LOCAL FOLDER
------------------------------------------------------------------------
Create a new folder named monument-sample-output next to (not inside) the approved source. Copy or write ONLY the files below into it. Write plain, matter-of-fact English.

  README.md - what the asset is, possible buyer uses, what this sample does and does not contain, and the sample mode.
  ASSET_OVERVIEW.md - asset type, source types, counts (each marked measured / estimated / owner_asserted / unknown, with how it was obtained), formats, languages, coverage, how the asset was created if known, and known limitations.
  DATA_DICTIONARY.md - fields, file types or components and what they mean. Describe fields; do not paste real example values. Write "not applicable" where something does not apply.
  RIGHTS_AND_RESTRICTIONS.md - what I state about ownership and permission to license or use for AI training, any third-party content or licenses, customer or contract restrictions, and open questions. State clearly that these are owner statements that Monument has not verified.
  PRIVACY_AND_REVIEW.md - what was deliberately left out, the sanitization applied, which checks actually ran (including the offline checker's result) and what they cannot prove, possible direct or indirect identifiers, and remaining risks.
  REVIEW_REQUIRED.md - a short prioritized checklist of decisions and approvals still needed, including the decision to share this package with Monument Exchange.
  sample_manifest.json - described below.
  CHECKSUMS.sha256 - described below.
  SAMPLES/SYNTHETIC/ - mode B only: labelled fictional examples.
  SAMPLES/AUTHORIZED/ - mode B only: examples the owner is demonstrably authorized to share, each listed in "authorized_examples" with its basis.
  SAMPLES/EXCERPTS/ - mode C only: the approved, sanitized real excerpts.
Omit SAMPLES/ entirely for metadata_only. Do not add any other files: no hidden files, no .git or dependency folders, no nested ZIP or archive files, no shortcuts or symbolic links, no executables.

sample_manifest.json must be valid JSON with exactly this structure (use null or "unknown" rather than guessing):

{
  "schema_version": "3.0",
  "prompt_version": "monument-sample-prompt-3.0.0",
  "package_id": "msp-YYYYMMDD-xxxxxx (today's date and 6 random letters or digits)",
  "asset_type": "codebase | structured_data | documents_text | agent_trajectories | rl_environment | ai_training_data | multimodal | other",
  "sample_mode": "metadata_only | synthetic_or_authorized_examples | real_excerpts",
  "generated_at": "ISO 8601 UTC time, or null",
  "generated_with": "name of the assistant or tool, or null",
  "processing": {
    "environment": "offline_local | approved_ai_assistant | manual",
    "assistant_cloud_connected": null,
    "organization_approval": "confirmed | not_required | unknown"
  },
  "intended_recipient": "Monument Exchange",
  "source_summary": {
    "description": "one or two sentences with no confidential detail",
    "population_count": null,
    "population_count_status": "measured | estimated | owner_asserted | unknown",
    "sampling_method": "how the examples were chosen, or none (metadata only)"
  },
  "rights": {
    "ownership": "owner_asserted | unknown | disputed",
    "third_party_content": "none_known | present | unknown",
    "ai_training_permission": "owner_asserted | not_permitted | unknown",
    "verified_by_monument": false
  },
  "privacy": {
    "personal_data_in_sample": "none_known | present_reviewed | unknown",
    "human_review_complete": false
  },
  "disclosure_authorization": null,
  "authorized_examples": [],
  "local_checker": null,
  "rl_environment": null,
  "local_checks": [
    { "name": "name of a check you ran", "result": "pass | warn | fail | not_run", "limits": "what this check cannot establish" }
  ],
  "owner_approval": {
    "status": "pending",
    "approved_by_role": null,
    "approved_at": null,
    "scope": null
  },
  "files": [
    { "path": "README.md", "sha256": "64 lowercase hex characters", "bytes": 1234, "role": "documentation | synthetic_example | authorized_example | real_excerpt" }
  ],
  "limitations": ["plain-language limitations of this sample"]
}

Notes on the manifest:
  - Pick one value where alternatives are separated by " | ". "assistant_cloud_connected" is true, false or null (unknown).
  - Mode C only: set "disclosure_authorization" to { "approved_by_role": "the approver's role, not their name", "approved_at": "ISO 8601 UTC", "basis": "why sharing these excerpts with Monument is permitted", "scope": "exactly what was approved", "excerpt_count": <number of files in SAMPLES/EXCERPTS/>, "sanitization_applied": ["each step, e.g. removed direct identifiers"], "re_checked_after_sanitization": true }. Set "privacy.human_review_complete" to true only after a person has reviewed every excerpt.
  - Mode B: list every file in SAMPLES/AUTHORIZED/ in "authorized_examples" as { "path": "...", "authorization_basis": "why the owner may share it" }. If you include synthetic examples, add a limitation saying they are synthetic and say nothing about the quality or characteristics of the real data.
  - Roles: files in SAMPLES/SYNTHETIC/ are "synthetic_example", in SAMPLES/AUTHORIZED/ "authorized_example", in SAMPLES/EXCERPTS/ "real_excerpt"; root documents are "documentation".
  - "files" lists every file in the package except sample_manifest.json and CHECKSUMS.sha256, with its exact relative path (forward slashes), byte size and SHA-256.
  - "local_checker" stays null until you run the offline checker (STEP 5); then set it to { "tool": "monument-sample-check", "version": "<the checker's version>", "candidate_status": "<the status it reported>" }.
  - For rl_environment assets, set "rl_environment" to { "runnable_claimed": true or false, "execution_evidence_included": true or false }. Otherwise leave it null.
  - Leave owner_approval as "pending" until STEP 9.

CHECKSUMS.sha256 has one line per file in the package except itself (so it includes sample_manifest.json), in the form:
  <64 lowercase hex sha256><two spaces><relative path>
Paths are relative to the monument-sample-output folder and use forward slashes. Compute sizes and checksums with a local tool: the offline checker prints them ("python3 monument_sample_check.py hashes monument-sample-output"); otherwise shasum -a 256, sha256sum, Get-FileHash, certutil or Python's hashlib. If no tool is available, say so instead of guessing.

------------------------------------------------------------------------
STEP 5 - RUN PRIVACY AND SECRET DETECTION ON THE CANDIDATE FILES
------------------------------------------------------------------------
On monument-sample-output only (never on my whole source):
  1. If I allowed it in STEP 1, I download monument_sample_check.py from the Monument sample page (Get the offline checker) and compare its SHA-256 with the one shown on that page. Then run, on this computer:
       python3 monument_sample_check.py check-folder monument-sample-output
     If I listed anything that must never be included (STEP 1, question 7), put those terms in a local text file OUTSIDE monument-sample-output (one per line) and add: --terms <that file>. The checker never prints the terms; never put that file in the package.
     It reports one of: CHECKS_PASSED (no findings by its rules in what it could scan), SENSITIVE_DATA_FOUND, REVIEW_REQUIRED, SCAN_INCOMPLETE, UNSUPPORTED_CONTENT, NEEDS_CORRECTION (package format) or SECURITY_HOLD (unsafe files). It never prints the detected values.
  2. If the checker is not allowed or cannot run, say so plainly, keep "local_checker" null and run what local checks you can (secret patterns such as private key blocks, "api_key=", "password=", tokens, cloud keys and connection strings; emails, phone numbers and ID-like numbers; hidden files; file names) - and record each one in local_checks with its limits. Never claim the offline checker ran when it did not.
  3. Also confirm: all required files are present and nothing extra is there; sample_manifest.json is valid; every listed file matches its size and SHA-256; CHECKSUMS.sha256 matches; metadata_only packages contain no SAMPLES/ folder; samples are only in the three SAMPLES/ folders; every number is marked measured / estimated / owner_asserted / unknown.

------------------------------------------------------------------------
STEP 6 - REPORT FINDINGS WITHOUT SECRET VALUES
------------------------------------------------------------------------
Give me a findings report: for each finding, its status or severity, the kind (for example "possible credential", "customer records", "internal hostname"), the file path and line number if known, and what you propose to do. Do not quote, paste or paraphrase any detected value. Do not send findings anywhere.

------------------------------------------------------------------------
STEP 7 - REMOVE OR REPLACE, THEN CHECK AGAIN
------------------------------------------------------------------------
Remove or replace every finding (delete the file, replace real values with placeholders, use synthetic rows, generalise rare combinations, strip document properties, rename revealing files). Rotating any real secret you found is my job: remind me. Then recompute the manifest and checksums and repeat STEP 5 until the checker reports CHECKS_PASSED, or until the only remaining items are REVIEW_REQUIRED / UNSUPPORTED_CONTENT findings that a named role in my organization has reviewed and explicitly accepted (record that in PRIVACY_AND_REVIEW.md). SENSITIVE_DATA_FOUND, SCAN_INCOMPLETE, NEEDS_CORRECTION and SECURITY_HOLD must be fixed - never accepted.

EXCEPTIONAL CASES: if you or I believe real personal, regulated or confidential data genuinely needs to be shared, STOP. Do not include it. Tell me to contact my Monument contact first: such cases need Monument's privacy specialist and a separate legal authorization before anything is shared, and may still be refused.

------------------------------------------------------------------------
STEP 8 - SHOW ME EXACTLY WHAT WILL BE INCLUDED
------------------------------------------------------------------------
Show me a plain-English disclosure review:
  - the asset type and sample mode (A, B or C)
  - the exact list of files that would be shared, with sizes and roles
  - for mode C: each excerpt, its source, the sanitization applied and the disclosure authorization
  - anything sensitive you noticed and what you did about it
  - rights questions that are still open
  - the intended recipient: Monument Exchange
  - the offline checker's result (or that it did not run) and the limits of every check

------------------------------------------------------------------------
STEP 9 - STOP FOR MY EXPLICIT APPROVAL
------------------------------------------------------------------------
Ask: "Do you approve sharing exactly these files with Monument Exchange for evaluation? (yes / no / change something)"
Do not create a ZIP until I answer yes. If I say no, stop and do not package anything. If I ask for changes, make them, repeat STEPS 5 to 8, and ask again.
After I say yes: update owner_approval in sample_manifest.json ("status": "approved", "approved_by_role": my role or title (not my name), "approved_at": ISO 8601 UTC, "scope": one sentence describing what I approved for Monument Exchange), recompute CHECKSUMS.sha256 and run STEP 5 once more.

------------------------------------------------------------------------
STEP 10 - CREATE THE ZIP LOCALLY
------------------------------------------------------------------------
Create monument-sample-output.zip next to the folder, containing only the monument-sample-output folder, using a local tool. For example:
  macOS / Linux:   cd <parent folder> && zip -r -X monument-sample-output.zip monument-sample-output -x "*.DS_Store" "__MACOSX/*"
  Python 3:        python3 -c "import shutil; shutil.make_archive('monument-sample-output', 'zip', root_dir='.', base_dir='monument-sample-output')"
  Windows 10/11:   tar.exe -a -c -f monument-sample-output.zip monument-sample-output
Do not password-protect or encrypt the ZIP, do not add a ZIP comment, and do not put other archives inside it.
If you cannot create a ZIP yourself, do not claim you did. Give me these manual steps instead:
  macOS: right-click the monument-sample-output folder and choose "Compress".
  Windows: right-click the folder, choose "Send to" > "Compressed (zipped) folder" (or "Compress to ZIP file").
  Linux: run the zip command above in a terminal.

------------------------------------------------------------------------
STEP 11 - INSPECT THE FINAL ZIP ITSELF
------------------------------------------------------------------------
Check the ZIP that will actually be shared, not just the folder: every entry, its name, its metadata and its contents.
  python3 monument_sample_check.py check-zip monument-sample-output.zip   (add --terms <file> if you used it in STEP 5)
Report the ZIP as "locally check-complete" ONLY if every mandatory check completed and the checker reported CHECKS_PASSED for the ZIP. If any check could not run, did not finish, or reported anything else, say "not locally check-complete" and why. Then tell me: the full local path of monument-sample-output.zip, its size, its SHA-256 (if a tool is available), the asset type and sample mode, and any warnings. Never describe a clean result as proof of anonymization or legal clearance.

------------------------------------------------------------------------
STEP 12 - NEVER UPLOAD IT YOURSELF
------------------------------------------------------------------------
Do not upload, email or send the ZIP anywhere, and do not open a browser to do so. Do not send the findings or the checker's report anywhere either. Tell me to go to https://monumentexchange.com/create-sample, use my secure Monument invitation link, and submit the ZIP in the "Submit your sample" section. Remind me that Monument checks every ZIP independently, that possible sensitive or uncheckable content is held for a restricted Monument reviewer, that real excerpts are always reviewed by a Monument privacy reviewer before anyone else, and that a received or checked sample is not approved for any buyer.

==================== METADATA-ONLY QUESTIONNAIRE ====================
(Use this when file access is not approved. Answer in plain text and send it to your Monument contact.)

  1. What is the asset, in one or two sentences?
  2. Asset type: codebase, structured_data, documents_text, agent_trajectories, rl_environment, ai_training_data, multimodal, or other.
  3. Roughly how large is it (files, rows, hours, repositories)? Is that measured or an estimate?
  4. What formats and languages does it use?
  5. How was it created or collected, and over what period?
  6. Who owns it, and do you believe you can license it (including for AI training)? Are there third-party, customer or employee restrictions?
  7. Does it contain personal data, customer data or confidential business information?
  8. What would a buyer most want to know before evaluating it?
  9. Which approved tool or process could later be used to prepare a sample, if any?