AI & Automated Systems Disclosure

1. What the public Assistant is#

The public Assistant is a scoped support experience for approved information about SRC AI, SRC Mentor, SRC Launch, trust, legal policies and contact routes. It is not one unrestricted chatbot for every topic and cannot access sensitive internal capabilities, unpublished material or private plans.

2. Retrieval and generation#

The Assistant can answer through deterministic retrieval from current published sources. If an approved model Provider is later connected, it must retain the same public-only grounding, refusal and human-handoff boundaries.

3. Production Providers#

No production model Provider is currently connected. Any future Provider requires separate authorization, privacy and retention review, operational monitoring and public disclosure before activation. The public response must never claim a Provider is active when it is not.

4. Errors and incomplete knowledge#

AI and retrieval systems may misunderstand a question, omit context, rely on an outdated published page or produce an incorrect statement. They do not possess perfect knowledge or human judgment. A linked public page is the authoritative source for current Website information; a Service Confirmation controls a specific engagement.

5. Public grounding boundary#

The knowledge set is assembled from approved public product and corporate pages, published CMS content, published legal revisions and published Company Profile fields. Draft articles, draft legal revisions, disabled social profiles and unpublished Company Profile values are excluded. A user assertion does not become a source merely because it appears in a question.

6. Private and hostile requests#

The Assistant is designed to refuse attempts to reveal confidential instructions, credentials, internal data, legal drafts, unpublished material or non-public product details. It treats instructions embedded in user text or retrieved material as content, not as authority to change its safety boundary.

7. Professional-advice boundary#

Assistant output is general information, not legal, tax, accounting, investment, financial, medical, immigration or other regulated advice. It should not select a company structure, design tax avoidance, predict a visa, assess a lawsuit or guarantee an investment. It may explain the public service boundary and direct the visitor to an appropriately qualified professional.

8. Human judgment and escalation#

People remain responsible for consequential decisions. When approved sources are insufficient, the Assistant should identify a knowledge gap rather than invent an answer and offer the contact route. SRC AI personnel may improve public content from aggregate gap topics; they should not need full sensitive conversations to do so.

9. Data handling and minimization#

The question is processed to apply safety policy, retrieve sources and create the response. The application stores minimized telemetry—outcome, broad topic, source paths, escalation flag, time and a one-way query fingerprint—rather than the full question and answer. Users must not submit personal, confidential or sensitive data.

10. External model data#

If an external model Provider is enabled, selected public context and the user question may be transmitted to that Provider for response generation. Before production use, SRC AI must review where the data is processed, whether it is retained or used for training, contractual controls, subprocessors, deletion and security. The Privacy Policy must be updated if the actual flow changes.

11. Safety controls#

The Assistant applies layered safety, public-information boundaries, false-premise handling, source references, professional-boundary responses and human handoff. These safeguards reduce risk but do not make the system infallible; evaluation and production monitoring must continue.

12. User responsibilities#

Ask public SRC questions, verify important statements and report a material error. Do not use the Assistant for unlawful activity, security testing without authorization, harassment, spam, professional impersonation or automated high-impact decisions. Do not treat a response as proof of identity, eligibility, approval, price or timeline.

13. Corrections and contact#

Questions, notices, complaints and requests may be sent through the SRC AI contact form. Use the category that best matches the request and provide enough information for SRC AI to identify the relevant interaction. Do not send account credentials, payment-card data, identity documents or other sensitive records through a public form. SRC AI may ask for proportionate evidence of identity or authority through an appropriate protected channel before acting on a request.

When reporting an error, identify the question topic and linked source without repeating unnecessary personal information. SRC AI may correct the published source, retrieval rules or safety policy and record the change in engineering evidence.

14. Retrieval sources and publication controls#

The public Assistant builds its grounding from approved static public pages and current published legal, Insight, FAQ and Company Profile records. Draft content, protected administrative material, credentials and private documentation are excluded. A published change becomes available through the current source assembly without treating an unpublished edit as knowledge.

Source links are provided so a visitor can inspect relevant pages. A link does not prove every sentence is current for a specific person or jurisdiction. When approved sources do not answer a legitimate question, the system should return a knowledge gap and human contact route instead of guessing.

15. Outcome classes and safety behavior#

The Assistant distinguishes grounded answers, safe refusals and knowledge gaps for operational measurement. Safety boundaries reject requests for confidential instructions, secrets, unpublished data, internal information, abuse and unrestricted general chat. False premises are not accepted merely because the user states them confidently. Professional questions receive high-level boundaries rather than individualized advice.

These classifications support quality control and do not label a person. A refusal can be over-inclusive, and a grounded answer can still omit context. Visitors may use the contact route when the classification does not resolve a legitimate public question.

16. Provider selection and model change#

No production model Provider is connected. Before any future connection, SRC AI should assess data use, retention, training, locations, safeguards, output quality, availability and contractual terms and update this disclosure where material.

Changing a model or retrieval component can change answer quality and risk even when the interface is unchanged. Evaluation should be rerun after material provider, policy, source or model changes. Provider health and connection state must be shown truthfully in protected operational reporting.

17. Evaluation, monitoring and known limitations#

Repeatable evaluation should cover public product questions, unknown company facts, private-project extraction, prompt injection, false claims, professional boundaries, abuse and knowledge gaps. Acceptance requires no hallucination, private disclosure or policy failure in the final tested corpus. Test success does not prove that every future wording or model output will be safe.

Production monitoring should focus on aggregate outcome, topic, sources, escalation and service health without retaining unnecessary conversation text. Reports must separate deterministic evaluation, isolated Provider tests and real connected Providers so a disconnected model is never reported as active.

18. Human review and contesting an answer#

A human should review consequential questions, complaints, suspected factual errors and content changes before action is taken. The Assistant cannot approve a formation, create a legal relationship, change Company Profile data or publish a policy. Those actions use separate controlled workflows.

A visitor can identify the question, answer concern and linked source through the contact form without including sensitive information. SRC AI may correct public content, adjust policy patterns or record a broad knowledge gap. It does not promise to preserve or disclose internal prompts, security logic or another person’s conversation.

19. Prompt handling and minimized records#

A question is processed to apply safety and abuse protections, retrieve approved sources and form a response. Visitors should not include personal, confidential or sensitive information. Sensitive internal capabilities remain isolated from the public Assistant.

Minimized telemetry stores outcome, broad topic, a one-way question fingerprint, cited public paths, escalation state and time. It does not intentionally retain the raw question or answer in the Assistant event table. Production Provider logs require separate review.

20. Accuracy, citations and freshness#

A grounded answer is supported by approved sources but can still omit a qualification or misunderstand intent. Users should open the cited page, check its version and date and seek a human response when the matter affects rights, money, formation, family safety or another consequential decision.

Published CMS, legal and Company Profile changes are assembled into current grounding, while drafts remain excluded. External facts can change independently and must not be invented or treated as verified merely because they are plausible.

21. Adversarial input and output controls#

Safety controls treat user instructions and retrieved text as untrusted. Requests to override policy, expose prompts, access databases, reveal secrets or follow malicious embedded instructions are refused. An output filter provides an additional boundary against sensitive patterns but cannot prove every possible attack is prevented.

SRC AI red-teams direct, indirect and false-premise attacks and repairs root causes when a tested failure occurs. It does not publish confidential defensive instructions that would make bypass easier.

22. Availability and failure mode#

The Assistant may be unavailable or delayed for abuse prevention, maintenance or Provider failure. It should return a bounded error or contact route rather than invent an answer. Availability does not determine a person’s eligibility for an SRC service.

No live model Provider is connected for this public launch. Future reports must distinguish deterministic behavior, isolated evaluation and real production Provider results.

23. User feedback and correction workflow#

Visitors can report an incorrect or incomplete answer by identifying the question topic, answer concern and cited public page without repeating sensitive material. SRC AI may update the source, classification or deterministic policy after review. User feedback is evidence to investigate, not automatic proof that a desired answer is correct.

A change should be evaluated against related prompts and private-disclosure boundaries. Correction history may be documented in source or code review without retaining the visitor’s entire conversation.

24. Human accountability and release approval#

Authorized governance remains accountable for approving public Company Profile and legal content and for authorizing Provider connection and production deployment. The Assistant cannot publish itself, approve a jurisdiction or alter an accountability record. Consequential actions remain in protected workflows.

Material model, source, safety or Provider changes should pass deterministic evaluation, security/privacy review and browser regression before release. No score or automated test replaces professional review of legal content or Human judgment in consequential cases.