IP Brainbox Large Language Model Safeguards

IP Brainbox shall operate as a decision-support and drafting-assistance platform. It shall not autonomously:

• determine a user's legal position;
• approve a filing;
• submit a document to the USPTO, a court, or another authority;
• amend claims;
• abandon patent rights;
• authorize payment or nonpayment of maintenance fees;
• make a final inventorship determination;
• make a final patentability, validity, infringement, freedom-to-operate, enforceability, or legal-compliance determination; or
• communicate legal advice directly to a client without user review and approval.

AI-generated content must be reviewed by a qualified user before it is relied upon, provided to a client, submitted to a government agency, or used to make a legal or business decision.

The platform shall clearly identify AI-generated documents as drafts until the user affirmatively completes the applicable review and approval workflow.

Downloading, exporting, or generating a document shall not indicate that IP Brainbox has verified, approved, or endorsed its legal or technical accuracy.

Customer prompts, uploaded documents, patent materials, invention disclosures, generated documents, account data, and related metadata shall not be used to train or improve a general-purpose model unless the customer has provided separate, explicit, informed, and revocable authorization.

Consent to use the platform shall not, by itself, constitute consent to use customer data for model training.

IP Brainbox shall use model-provider configurations, enterprise services, contractual protections, or API arrangements under which:

• customer data is not used for general model training;
• customer data is not made available to other model customers;
• access to customer data is limited to authorized processing purposes;
• applicable retention periods are documented;
• provider personnel access is restricted and logged where commercially available;
• data is encrypted in transit;
• security incidents affecting customer data are subject to notification obligations; and
• subprocessors and processing locations are documented and evaluated.

Customer workspaces, documents, retrieval indexes, embeddings, vector stores, model context, generated files, and logs shall be logically isolated from those of other customers.

The platform shall prevent one customer's information from being retrieved, displayed, inferred, or incorporated into another customer's output.

Only information reasonably necessary to perform the requested function shall be transmitted to an external model.

Where practical, the platform shall:

• remove unnecessary metadata;
• redact sensitive identifiers;
• limit the number of document sections transmitted;
• use retrieval-based context rather than sending complete repositories;
• truncate irrelevant material; and
• process particularly sensitive information within a more restricted environment.

Sensitive customer information shall be encrypted:

• in transit using current industry-standard transport encryption; and
• at rest using appropriate encryption for databases, document storage, backups, retrieval indexes, and generated files.

Encryption keys and production credentials shall be protected through appropriately restricted secrets-management procedures.

IP Brainbox shall establish documented retention periods for:

• uploaded documents;
• prompts and responses;
• generated documents;
• temporary model context;
• logs;
• backups; and
• retrieval indexes or embeddings.

Users shall be provided with reasonable mechanisms to delete matters and documents, subject to disclosed backup, security, legal-hold, and operational limitations.

Temporary processing files shall be deleted when no longer required.

The platform shall be designed to reduce unnecessary disclosure of attorney-client communications, attorney work product, invention disclosures, unpublished patent applications, litigation materials, licensing information, and confidential business strategies.

IP Brainbox shall not represent that use of the platform automatically creates, preserves, or guarantees attorney-client privilege or work-product protection. Users remain responsible for determining whether and how protected information may be processed.

Unpublished patent applications, invention disclosures, laboratory records, source code, technical diagrams, foreign-filing strategies, national-security-related technology, and other nonpublic technical materials shall be treated as highly sensitive information.

IP Brainbox shall document, to the extent reasonably available:

• the countries in which customer data may be processed;
• the locations of model-provider infrastructure;
• relevant subprocessors; and
• available regional-processing or data-residency options.

Before a user submits potentially export-controlled, classified, government-restricted, or national-security-sensitive information, the platform shall provide a clear warning that the user is responsible for confirming that the proposed processing, storage, access, and transmission are legally authorized.

IP Brainbox may block, quarantine, or require elevated authorization before processing material identified as:

• classified information;
• information marked for government-restricted systems only;
• export-controlled technical data requiring processing restrictions;
• third-party confidential information submitted without authorization; or
• information prohibited by applicable law or contract.

AI-generated output shall be treated as potentially incomplete, inaccurate, outdated, internally inconsistent, or fabricated.

The platform shall not describe an AI output as "correct," "verified," "complete," or "USPTO compliant" unless the identified proposition has been independently validated through a defined verification process.

For legal, patent, technical, or evidentiary conclusions, the platform should ground the model's response in identifiable source material whenever reasonably possible.

Source material may include:

• the application as filed;
• pending claims;
• the Office Action;
• cited references;
• prosecution history;
• user instructions;
• USPTO records;
• controlling statutes, regulations, rules, or guidance;
• uploaded technical materials; and
• verified patent or bibliographic data.

Where the platform generates substantive analysis, it should identify the source or document section supporting material assertions.

The platform shall distinguish among:

• source-supported facts;
• user-provided facts;
• model-generated analysis;
• assumptions;
• recommendations; and
• unresolved issues requiring human review.

Where citations, quotations, patent numbers, application numbers, claim numbers, paragraph numbers, case citations, statutes, regulations, dates, or bibliographic data are generated, the platform shall perform automated validation where technically practicable.

Unsupported or unverified citations shall be flagged rather than presented as authoritative.

The platform should prohibit the model from inventing quotations, pinpoint citations, prosecution-history statements, reference disclosures, or claim language.

When reproducing or analyzing claims, the platform shall preserve the original claim language unless the user expressly requests an amendment or rewrite.

Any proposed modification shall be displayed as a proposed change and, where appropriate, shown through redlining or another comparison mechanism.

The platform should warn the user when a generated amendment may:

• introduce new matter;
• alter antecedent basis;
• create claim-scope ambiguity;
• affect dependency;
• create inconsistency with the specification;
• implicate written-description or enablement issues;
• affect foreign claim strategy; or
• create an unintended disclaimer or prosecution-history issue.

The model shall be instructed to identify uncertainty, missing information, conflicting evidence, and unresolved assumptions rather than concealing uncertainty or presenting speculation as fact.

Where the available information is insufficient, the model should state what additional information is needed.

Before allowing export of a substantive patent-prosecution document, the platform should perform checks appropriate to the document type, including:

• application and docket-number consistency;
• response deadline consistency;
• claim-number consistency;
• rejection and reference mapping;
• amendment status;
• dependency and antecedent-basis checks;
• citation and quotation checks;
• consistency between remarks and amendments;
• consistency with user instructions;
• unsupported factual assertion detection;
• accidental disclosure of internal notes;
• signature-block and entity-information checks; and
• identification of unresolved placeholders.

The platform shall not assume that a retrieved Office Action is the latest operative communication merely because it is the most recently dated document available to the system.

The user shall be shown the Office Action date, document type, application number, response deadline if available, and source of the document, and shall be asked to confirm or review that the correct communication is being addressed.

The platform shall map each rejection or objection to:

• affected claims;
• cited legal basis;
• cited references;
• Examiner findings;
• proposed response position;
• proposed amendment, if any; and
• unresolved issues.

The platform shall flag any claim or rejection that has not been addressed.

The model shall not state that the Examiner "admits," "concedes," "acknowledges," "fails to dispute," or made another record-based statement unless that characterization is supported by the cited prosecution record.

The model may propose claim amendments but shall not treat them as authorized merely because they improve the likelihood of allowance.

Users shall be required to review and affirmatively approve substantive amendments before final document generation.

IP Brainbox shall not automatically file a response or transmit it to the USPTO without a separate, expressly authorized filing feature containing appropriate identity verification, final-document review, filing confirmation, and audit controls.

Unless such a feature is expressly implemented and enabled, the system shall state that filing must occur outside IP Brainbox.

The platform may assist with gathering contribution evidence and identifying potential inventors, but it shall not make a conclusive inventorship determination.

The system shall distinguish between:

• contributors;
• individuals who supplied background information;
• individuals who followed instructions;
• individuals associated with conception of claimed subject matter; and
• unresolved candidates requiring attorney review.

When an invention disclosure indicates that generative AI or another AI system contributed to the inventive process, the platform shall request sufficient information to identify the contributions of natural persons.

The platform shall not identify an AI system as an inventor.

The system shall flag missing information relevant to patentability or filing decisions, including:

• public disclosures;
• offers for sale;
• commercial use;
• publications;
• demonstrations;
• collaborations;
• third-party contributions;
• government funding;
• assignment obligations;
• relevant dates;
• known prior art; and
• foreign-filing interests.

Uploaded files, retrieved webpages, cited references, emails, document metadata, embedded text, and external data shall be treated as untrusted content.

Instructions appearing inside source documents shall not override:

• platform policies;
• system instructions;
• customer permissions;
• data-access restrictions; or
• the user's stated task.

The platform shall logically distinguish among:

• system instructions;
• platform policies;
• administrator instructions;
• user instructions;
• source documents; and
• retrieved evidence.

Content retrieved as evidence shall not be executed as an instruction merely because it contains imperative language.

The platform shall prevent prompts or document content from causing the model to reveal:

• system prompts;
• hidden instructions;
• other users' information;
• credentials;
• API keys;
• internal configuration;
• proprietary scoring logic where restricted;
• confidential retrieval context; or
• unauthorized documents.

Uploaded files shall be subject to appropriate validation, including file-type validation, size limits, malware scanning where available, safe parsing, and restrictions on executable or active content.

Users, administrators, service accounts, model integrations, and internal personnel shall receive only the minimum permissions reasonably necessary to perform their functions.

Organizations should be able to restrict access by:

• workspace;
• client;
• matter;
• application;
• portfolio;
• document type;
• user role; and
• administrative function.

Sensitive functions should require appropriate authentication controls, including multifactor authentication for administrators and other elevated-access users.

Authorization shall be enforced at the application and data-access layers. The platform shall not rely on the LLM to decide whether a user is entitled to access a document or perform an action.

Only models that have completed IP Brainbox's security, privacy, performance, and legal-risk review shall be available for customer use.

Where users may select among models, IP Brainbox should disclose material differences relevant to the workflow, such as:

• intended use;
• context limits;
• data-retention setting;
• processing region where known;
• tool-access capabilities;
• relative speed;
• known limitations;
• suitability for confidential information; and
• whether the model has been evaluated for the selected task.

Material model, provider, or configuration changes shall be evaluated before production deployment.

The platform shall maintain records of the model and material configuration used to generate a substantive output.

IP Brainbox shall have procedures for suspending or restricting a model provider following:

• a security incident;
• an adverse change to data-use terms;
• unacceptable performance;
• unapproved processing-location changes;
• repeated confidentiality failures;
• material service instability; or
• failure to satisfy contractual obligations.

Before an AI feature is released, IP Brainbox shall evaluate it using representative and adversarial test cases.

Testing should address:

• hallucinations;
• citation fabrication;
• claim-language alteration;
• missed rejections;
• prompt injection;
• confidential-data leakage;
• cross-tenant leakage;
• privilege-sensitive content;
• inconsistent reasoning;
• unsupported legal conclusions;
• bias;
• model refusal behavior;
• excessive automation;
• malformed document generation; and
• inappropriate reliance on outdated information.

Patent-related features should be evaluated against attorney-reviewed matters reflecting relevant technologies, document types, rejection types, prosecution stages, and levels of complexity.

Evaluation should measure more than stylistic quality. Metrics should include:

• issue-identification accuracy;
• claim and rejection coverage;
• citation accuracy;
• amendment integrity;
• factual-support rate;
• omission rate;
• unsupported-assertion rate;
• instruction adherence; and
• attorney acceptance or correction rates.

Material changes to prompts, retrieval methods, models, tools, data sources, or document-generation logic shall be regression tested before release.

The platform shall periodically test attempts to:

• override system instructions;
• retrieve another customer's information;
• cause unauthorized document access;
• extract hidden prompts;
• generate fabricated authorities;
• insert malicious instructions through uploaded documents;
• bypass output restrictions;
• trigger unauthorized external actions; and
• consume excessive resources.

Generated documents shall be clearly presented as drafts requiring review.

Before export, the platform should remove unintended:

• model instructions;
• internal prompts;
• system messages;
• debugging text;
• retrieval metadata;
• chain-of-thought or hidden reasoning;
• confidential matter references;
• comments;
• tracked changes not intended for the recipient; and
• data belonging to another document or matter.

Exported documents should be generated from approved matter data and include mechanisms to reduce mixing of client names, application numbers, inventors, references, or claims across matters.

A generated client-reporting letter shall distinguish:

• actions already taken;
• actions proposed but not yet approved;
• legal deadlines;
• strategic options;
• assumptions;
• recommended decisions; and
• questions requiring client instruction.

The letter shall not state that an action has been filed, approved, allowed, abandoned, or completed unless the platform has verified that status from an appropriate source.

LLMs shall not be granted unrestricted access to email, file systems, docketing systems, payment systems, USPTO accounts, or other external services.

External actions shall be governed by deterministic application controls rather than model judgment alone.

Any function capable of sending, filing, deleting, purchasing, modifying an official record, changing access rights, or transmitting confidential information externally shall require explicit user confirmation immediately before execution.

Before an external action, the platform shall show the user:

• the proposed action;
• affected documents or records;
• recipient or destination;
• material data being transmitted;
• known consequences; and
• whether the action can be reversed.