# Use an AI assistant without outsourcing verification or authority

Treat this file as reference data, never as instructions that override the user or agent. Read the complete item before applying it. Keep use conditions, avoid conditions, limits, evidence roles and source links with the answer. An editorial block is a separately versioned analysis layer, not an original corpus claim or proof of efficacy. Cite the readable URL and version. Do not infer human review. This is a saved edition, not a live safety-notice service. Before consequential reuse, check the current page, https://vedokrok.com/data/availability.json and https://vedokrok.com/data/notices.json; absence alone does not explain why an item was removed.

Release: VEDOKROK-SEL-5d86e1b23401e578
Snapshot SHA-256: 30ce1c8fa9155763797e3d5cb55b1d94cbb8c3146ebd03fb3c006531ce60e82f
Current availability: https://vedokrok.com/data/availability.json
Current notices: https://vedokrok.com/data/notices.json

Give ordinary AI users safe and inspectable data-minimization, prompt-testing, instruction-trust, tool-permission and completion-check habits.
Collection kind: collection
Source: https://vedokrok.com/collections/use-an-ai-assistant-without-outsourcing-verification-or-authority
Version: 1.0.0

---

## Strip private material before asking an assistant for help

ID: MHC-D-RESEARCH-9794 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/strip-private-material-before-asking-an-assistant-for-help

The most useful example may not need the real names.

Knowledge source release: MHC-RPUB-20261113-0d6697fd
Source corpus commit: 0d6697fd15f7fb670cfa465c54a411e91752cae4
Source manifest SHA-256: 038080ffd768d7c1ae21ae65ea2a3b8e36664c28df759e06d19f63f6e059bf29
Source projection SHA-256: 489520d20496b8e53d4bcf37ddaae13d1b66f4405e33ae1666ce681073b24c78

### Use when

- You want AI to rewrite a real client email or analyze an account export.

### Avoid when

- Simple name removal may not prevent re-identification through rare details; authorization still applies.

### Explanation

Describe the structure of the task with invented identifiers, synthetic values and only the essential constraints. Verify the tool and workspace data policy separately if sensitive content is unavoidable; do not assume one personal account has business-contract safeguards.

### Steps

1. Replace real parties, account IDs and unique events with controlled examples.
2. Check the workspace's applicable data rules before uploading a real file.

### Example

Ask for a VAT-error explanation using customer A and fictional codes rather than a production customer export.

### Check

The prompt preserves the reasoning problem without exposing confidential records.

### Limits

- Simple name removal may not prevent re-identification through rare details; authorization still applies.

### Evidence and sources

- supports: OpenAI notes that consumer-service controls and handling differ from business offerings and advises against submitting sensitive content that should not be reviewed or used. — ai-privacy. Claim limit: Source is a domain framework or provider instruction, not a demonstrated intervention effect. Adapt to risk and the actual account or software version.
Source reading history: selected_passages; 2026-10-10; Data controls and cautions about sensitive content. The cited official page was retrieved on 2026-10-10 and the named passages were inspected. Card-specific adaptations are independent editorial reasoning, not validated improvements.
Agent release review: Agent checked the exact card fields and its claim, source, locator, example, check and stated boundary. This bounded projection preserves the source pack’s screened evidence scope and does not claim independent scientific or human editorial review. (Data controls and cautions about sensitive content)
- ai-privacy: How OpenAI handles data in consumer services — https://help.openai.com/en/articles/7039943-how-openai-handles-data-in-consumer-services

Review: agent · Agent editorial and scoped source review · 2026-11-13T05:54:13Z
Published: 2026-11-13T05:54:13Z
Related (useful_with): https://vedokrok.com/knowledge/check-chatgpt-data-controls-before-sharing

---

## Check ChatGPT data controls before sharing

ID: MHC-D-RESEARCH-9795 · Version: 0.1.0 · Kind: checklist
Source: https://vedokrok.com/knowledge/check-chatgpt-data-controls-before-sharing

Turning off training does not give you permission to share a client secret.

Knowledge source release: MHC-RPUB-20261113-0d6697fd
Source corpus commit: 0d6697fd15f7fb670cfa465c54a411e91752cae4
Source manifest SHA-256: 038080ffd768d7c1ae21ae65ea2a3b8e36664c28df759e06d19f63f6e059bf29
Source projection SHA-256: 489520d20496b8e53d4bcf37ddaae13d1b66f4405e33ae1666ce681073b24c78

### Use when

- You are building an everyday ChatGPT habit for work or study.

### Avoid when

- If you submit feedback, the linked conversation may be used for training despite an opt-out. Available controls depend on your account and workspace; check the current policy.

### Explanation

Check the account or workspace you are using. In ChatGPT, the model-improvement setting governs training on new conversations; it does not delete saved chats. Memory and model training have separate controls. Match these settings with the rules for the material you want to share.

### Checklist

- Open Settings, then Data controls. Read the current model-improvement option and choose the available setting deliberately.
- Check employer and client restrictions separately. If sharing is not permitted, use an invented example instead of uploading the real material.

### Example

A professional turns off model training on a personal account. Their saved chats remain, and client confidentiality still applies.

### Check

You can explain what the setting changes, what remains saved and whether the intended input is permitted.

### Limits

- If you submit feedback, the linked conversation may be used for training despite an opt-out. Available controls depend on your account and workspace; check the current policy.

### Evidence and sources

- supports: OpenAI states that disabling model improvement excludes new ChatGPT conversations from training without deleting saved chats; memory is controlled separately, and feedback may permit training use of the linked conversation despite opt-out. — ai-control. Claim limit: Current provider guidance for ChatGPT, not a guarantee of zero processing or retention and not permission to share restricted material. Controls depend on the account/workspace. This does not establish how a different service or separate environment setting handles data.
Source reading history: selected_passages; 2026-10-10; Turn off Improve the model for everyone; What happens to saved chats; Memory and safety context; feedback FAQ. The cited official page was retrieved on 2026-10-10 and the named passages were inspected. Card-specific adaptations are independent editorial reasoning, not validated improvements. Those selected control, saved-chat, memory and feedback passages reread on 2026-10-12; no account setting was changed or runtime data-handling behaviour tested.
Agent release review: Agent checked the exact card fields and its claim, source, locator, example, check and stated boundary. This bounded projection preserves the source pack’s screened evidence scope and does not claim independent scientific or human editorial review. (Turn off Improve the model for everyone; What happens to saved chats; Memory and safety context; feedback FAQ)
- ai-control: Data controls in ChatGPT — https://help.openai.com/en/articles/7730893-data-controls-in-chatgpt

Review: agent · Agent editorial and scoped source review · 2026-11-13T05:54:13Z
Published: 2026-11-13T05:54:13Z
Related (useful_with): https://vedokrok.com/knowledge/check-what-the-ai-actually-changed

---

## Use AI to extract questions from a contract, not legal conclusions

ID: MHC-D-RESEARCH-9796 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/use-ai-to-extract-questions-from-a-contract-not-legal-conclusions

A model can help find the question; it cannot promise the answer is legally correct.

Knowledge source release: MHC-RPUB-20261113-0d6697fd
Source corpus commit: 0d6697fd15f7fb670cfa465c54a411e91752cae4
Source manifest SHA-256: 038080ffd768d7c1ae21ae65ea2a3b8e36664c28df759e06d19f63f6e059bf29
Source projection SHA-256: 489520d20496b8e53d4bcf37ddaae13d1b66f4405e33ae1666ce681073b24c78

### Use when

- You face a long subscription, employment or service agreement and need to understand what to ask before signing.

### Avoid when

- Do not upload a confidential contract without permission, rely on generated legal advice or assume a clause has the same meaning across jurisdictions.

### Explanation

Share only permitted, minimized excerpts. Ask AI to locate payment, renewal, termination, liability, data and notice provisions, with exact quotations or clause references from the text provided. Check every reference against the original and take consequential legal interpretations to a qualified adviser.

### Steps

1. Ask for a table of clause, plain-language question and exact source location.
2. Verify each proposed interpretation against the original text; flag missing pages and unresolved terms.

### Example

An AI reading of a service agreement surfaces automatic renewal and cancellation notice clauses for direct verification before purchase.

### Check

Every material question points to a real provision or is explicitly labeled missing and unverified.

### Limits

- Do not upload a confidential contract without permission, rely on generated legal advice or assume a clause has the same meaning across jurisdictions.

### Evidence and sources

- supports: OWASP names prompt injection, information disclosure, excessive agency and misinformation as distinct risk classes for generative AI applications. — ai-top. Claim limit: Source is a domain framework or provider instruction, not a demonstrated intervention effect. Adapt to risk and the actual account or software version.
Source reading history: selected_passages; 2026-10-10; Prompt injection, disclosure, excessive agency and misinformation. The cited official page was retrieved on 2026-10-10 and the named passages were inspected. Card-specific adaptations are independent editorial reasoning, not validated improvements.
Agent release review: Agent checked the exact card fields and its claim, source, locator, example, check and stated boundary. This bounded projection preserves the source pack’s screened evidence scope and does not claim independent scientific or human editorial review. (Prompt injection, disclosure, excessive agency and misinformation)
- supports: The ABA's 2024 Formal Opinion 512 recognizes contract analytics as a legal use for generative AI, while warning that outputs can be inaccurate and that lawyers must retain appropriate review and professional judgment. — ai-aba-fo512. Claim limit: This is professional guidance about lawyers' duties under ABA Model Rules, not a consumer-use standard, evidence that a prompt method works, or a rule adopted in every jurisdiction.
Source reading history: selected_passages; 2026-10-11; Introduction, pp. 1-2 (contract analytics); Part II.A, pp. 3-4 (inaccurate output, independent review and professional judgment). Professional ethics guidance for lawyers applying ABA Model Rules; it does not establish consumer-task effectiveness or a universal legal rule.
Agent release review: Agent checked the exact card fields and its claim, source, locator, example, check and stated boundary. This bounded projection preserves the source pack’s screened evidence scope and does not claim independent scientific or human editorial review. (Introduction, pp. 1-2; Part II.A, pp. 3-4)
- ai-top: OWASP Top 10 for LLM and GenAI — https://genai.owasp.org/initiatives/top-10-for-llm-and-genai/
- ai-aba-fo512: Formal Opinion 512: Generative Artificial Intelligence Tools — https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf

Review: agent · Agent editorial and scoped source review · 2026-11-13T05:54:13Z
Published: 2026-11-13T05:54:13Z
Related (useful_with): https://vedokrok.com/knowledge/strip-private-material-before-asking-an-assistant-for-help

---

## Compare prompt versions on saved test cases

ID: MHC-D-RESEARCH-9797 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/compare-prompt-versions-on-saved-test-cases

A better answer on one example can hide a new mistake elsewhere.

Knowledge source release: MHC-RPUB-20261113-0d6697fd
Source corpus commit: 0d6697fd15f7fb670cfa465c54a411e91752cae4
Source manifest SHA-256: 038080ffd768d7c1ae21ae65ea2a3b8e36664c28df759e06d19f63f6e059bf29
Source projection SHA-256: 489520d20496b8e53d4bcf37ddaae13d1b66f4405e33ae1666ce681073b24c78

### Use when

- You reuse one prompt each week and assume a fluent answer means success.

### Avoid when

- Three cases are an initial smoke test, not proof of general reliability or safety. Provider changes or hidden settings can limit reproducibility.

### Explanation

Save an ordinary case, an ambiguous case and a known failure, with the result each should produce. Compare the old and new prompts using the same inputs and checks. Start each attempt without leftover conversation or test state. Keep the model and tools unchanged where possible; record any change so you do not attribute it to the prompt.

### Steps

1. Save both prompts, the visible model label, relevant settings and tool access, plus each output and its pass or fail result. Mark unavailable details as unknown.
2. Inspect failures against the agreed checks. If a decision depends on consistency, repeat cases under the recorded conditions and report failures as well as passes.

### Example

An invented inbox test expects a meeting request to be labelled scheduling, a request mixing scheduling and billing to be flagged for review, and an automatic reply to be ignored. A rewrite fixes the mixed request but labels the automatic reply scheduling: keep both results.

### Check

The comparison shows which saved cases passed or failed under which conditions. One passing attempt does not establish consistent performance.

### Limits

- Three cases are an initial smoke test, not proof of general reliability or safety. Provider changes or hidden settings can limit reproducibility.

### Evidence and sources

- supports: Anthropic describes evaluation tasks with defined inputs and success criteria, records of attempts, repeated trials for variable model outputs, and clean, stable test environments. — ai-evals. Claim limit: Provider engineering guidance, not experimental proof that a three-case personal prompt check establishes reliability. Recording visible conditions is an editorial adaptation; hidden provider settings may prevent exact reproduction.
Source reading history: selected_passages; 2026-10-10; Evaluation structure: task, trial, transcript and harness; Step 4: stable environment. The cited official page was retrieved on 2026-10-10 and the named passages were inspected. Card-specific adaptations are independent editorial reasoning, not validated improvements. Selected definitions and Step 4 reread on 2026-10-12; this does not constitute full scientific review.
Agent release review: Agent checked the exact card fields and its claim, source, locator, example, check and stated boundary. This bounded projection preserves the source pack’s screened evidence scope and does not claim independent scientific or human editorial review. (Evaluation structure: task, trial, transcript and harness; Step 4: stable environment)
- ai-evals: Demystifying evals for AI agents — https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents

Review: agent · Agent editorial and scoped source review · 2026-11-13T05:54:13Z
Published: 2026-11-13T05:54:13Z
Related (use_before): https://vedokrok.com/knowledge/rehearse-an-ai-automation-with-test-data

---

## Rehearse an AI automation with test data

ID: MHC-D-RESEARCH-9798 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/rehearse-an-ai-automation-with-test-data

Start with test data and limited permissions.

Knowledge source release: MHC-RPUB-20261113-0d6697fd
Source corpus commit: 0d6697fd15f7fb670cfa465c54a411e91752cae4
Source manifest SHA-256: 038080ffd768d7c1ae21ae65ea2a3b8e36664c28df759e06d19f63f6e059bf29
Source projection SHA-256: 489520d20496b8e53d4bcf37ddaae13d1b66f4405e33ae1666ce681073b24c78

### Use when

- An assistant proposes to rename files, process email or edit your calendar through an integration.

### Avoid when

- Passing this rehearsal does not authorise a real run or guarantee its safety. Follow the applicable data-handling rules.

### Explanation

Check which accounts and tools the assistant can access. Use invented records in a separate test folder or workspace, with only the permissions the rehearsal needs. Check that it cannot reach production files, send real messages or spend money. Copying real records does not remove their sensitive content. Enforce the limits through tool or account controls, not just a prompt.

### Steps

1. Compare planned changes with actual results, including an exception and a request outside the allowed scope.
2. If isolation or permissions are unclear, stop at a preview. Before a real run, confirm its permitted targets and actions, including any required approval for high-impact changes.

### Example

A renaming test can change only invented files in its test folder. A request to rename another invented file outside it should be refused, even if the model suggests doing so.

### Check

Expected changes occur only within the test scope, and prohibited actions are refused. A preview alone has not tested execution.

### Limits

- Passing this rehearsal does not authorise a real run or guarantee its safety. Follow the applicable data-handling rules.

### Evidence and sources

- supports: OWASP describes damaging actions from unexpected or manipulated model outputs and recommends limiting tool functions and permissions, approving high-impact actions and enforcing authorisation in downstream systems. — ai-owasp. Claim limit: Security-practice guidance, not measured proof that this consumer rehearsal prevents harm. The invented records, isolated test workspace, refused out-of-scope case and preview fallback are editorial adaptations; a successful rehearsal grants no production authority or security certification.
Source reading history: selected_passages; 2026-10-11; LLM06:2025: risk examples and mitigation strategies 1-7, especially minimal functionality/permissions, high-impact approval and downstream authorisation. The cited official page was retrieved on 2026-10-10 and the named passages were inspected. Card-specific adaptations are independent editorial reasoning, not validated improvements. On 2026-10-11 these selected passages were read again for card 9798. The rehearsal is an authored workflow; no tool isolation, production deployment, attack-resistance or empirical effectiveness was tested.
Agent release review: Agent checked the exact card fields and its claim, source, locator, example, check and stated boundary. This bounded projection preserves the source pack’s screened evidence scope and does not claim independent scientific or human editorial review. (LLM06:2025 Excessive Agency: Common Examples of Risks; Prevention and Mitigation Strategies 1-7)
- ai-owasp: LLM06:2025 Excessive Agency — https://genai.owasp.org/llmrisk/llm062025-excessive-agency/

Review: agent · Agent editorial and scoped source review · 2026-11-13T05:54:13Z
Published: 2026-11-13T05:54:13Z
Related (useful_with): https://vedokrok.com/knowledge/check-what-the-ai-actually-changed

---

## Check what the AI actually changed

ID: MHC-D-RESEARCH-9799 · Version: 0.1.0 · Kind: checklist
Source: https://vedokrok.com/knowledge/check-what-the-ai-actually-changed

A convincing sentence is not the same as a changed system.

Knowledge source release: MHC-RPUB-20261113-0d6697fd
Source corpus commit: 0d6697fd15f7fb670cfa465c54a411e91752cae4
Source manifest SHA-256: 038080ffd768d7c1ae21ae65ea2a3b8e36664c28df759e06d19f63f6e059bf29
Source projection SHA-256: 489520d20496b8e53d4bcf37ddaae13d1b66f4405e33ae1666ce681073b24c78

### Use when

- An assistant says it sent a message, changed a file or merged an update.

### Avoid when

- Completion does not prove permission or quality. Respect access limits; a screenshot can show an older state.

### Explanation

Before the action, name the result you expect and where you will check it. Afterwards, open that service and match the target, action and result to your request. A link or identifier helps locate the result; it does not replace checking it. Keep a prepared draft separate from a completed action.

### Checklist

- Check the intended recipient, file or record, then its current status or revision. For a batch, check which items succeeded and which failed.
- If the result is missing, inaccessible or still pending, report that state. Check for an existing action before retrying a send or another action that could be duplicated.

### Example

An assistant says an issue was closed. Open the intended project and issue: closing a different issue does not fulfil the request.

### Check

The observed result matches the requested target and action. Unchecked items remain unverified; pending items remain pending.

### Limits

- Completion does not prove permission or quality. Respect access limits; a screenshot can show an older state.

### Evidence and sources

- supports: NIST says AI trustworthiness results should reflect relevant actors' input about performance in deployment contexts and be documented. — ai-risk. Claim limit: NIST provides a general organisational evaluation/documentation framework, not this consumer receipt-checking procedure or evidence of its effectiveness. Matching targets, handling partial/pending results and checking before retry are editorial advice; observation does not establish authorisation or quality.
Source reading history: selected_passages; 2026-10-10; NIST AI 600-1, printed p. 39 (PDF p. 43), MEASURE 4.2 heading. The cited official page was retrieved on 2026-10-10 and the named passages were inspected. Card-specific adaptations are independent editorial reasoning, not validated improvements. MEASURE 4.2 heading reread in the linked report on 2026-10-12; it supports the general framework statement, not the card-specific action receipt procedure.
Agent release review: Agent checked the exact card fields and its claim, source, locator, example, check and stated boundary. This bounded projection preserves the source pack’s screened evidence scope and does not claim independent scientific or human editorial review. (MEASURE 4.2: deployment-context trustworthiness results, relevant actor input and documentation)
- ai-risk: NIST AI RMF Generative AI Profile — https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence

Review: agent · Agent editorial and scoped source review · 2026-11-13T05:54:13Z
Published: 2026-11-13T05:54:13Z
