Acceptable Use Policy

1. The Good Citizen clause

You direct the work on greatVibe. You own the result. You are responsible for what runs.

That is the whole clause. The rest of this section is what it means in practice.

1.1 Why this clause exists

greatVibe runs your directed work on shared infrastructure. AWS provides the compute. Anthropic, OpenAI, and other model providers run the inference you connect. Other greatVibe customers share the same mesh nodes when they share the same plan.

That shared layer is not a bug. It is how a turn-based platform stays affordable. It does mean the cost of one customer abusing it falls on everyone else. We have to keep that shared infrastructure clean. So do you.

1.2 What you commit to

You commit to direct your sessions in a way that does not:

• Scan or probe networks you do not own.
• Reach private network ranges from any greatVibe mesh node. Private RFC1918 ranges and the cloud metadata range are out of bounds: 10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16, 169.254.0.0/16.
• Generate traffic designed to disrupt other services.
• Bypass access controls on systems you are not authorised to use.
• Send unsolicited messages at scale.
• Use greatVibe to generate scaled personalised content that deceives, intimidates, defrauds, or coordinates influence operations against identifiable individuals or organisations.
• Probe or exfiltrate cloud provider metadata.
• Run port scans, vulnerability scans, or reconnaissance against third parties.
• Direct the AI models you connect to produce output that breaches those providers' own acceptable use policies. Your obligations to Anthropic, OpenAI, or any other provider you connect are yours, not ours.

That is the practical list. The principle is the line at the top: you own the direction, so you own the result of the direction.

1.3 What we enforce, and how

We enforce the Good Citizen clause at three points. At each point, we monitor signals. We do not store or read your prompt content.

1. Mesh network boundary. Outbound traffic from your mesh node is filtered. The ranges and ports listed in section 1.2 are blocked at the network layer. Per-session outbound HTTP is rate-limited.
2. Provider refusal log. When the model provider you connect refuses one of your requests, we record the refusal signal: which provider, which user, which model, what kind of refusal. We do not record the prompt. Repeated refusals from one account trigger a review.
3. Tool-call audit. When a session runs commands, transfers files, or exports credentials, the action is recorded in an append-only log. Logs are retained for 12 months. We use them for incident response and for cooperation with cloud and model partners. We do not use them for marketing or analytics.

Your prompt content stays private. The same is true of the files you store in your workspace. We record what your session reached out and did. We do not record what you typed.

1.4 What we do not do

We do not scan prompts for keywords. We do not run sentiment analysis on your work. We do not infer your intent from your prompt. The providers you connect already filter their own output. We trust that boundary and stay out of yours.

Session payloads are processed on your own nodes. We don't store or retain the content of your sessions.

We do not auto-suspend on a single signal. A real person reviews every suspension decision. The Terms commit us to this. The AUP repeats it because it matters.

1.5 What happens when you breach

If we believe you have breached the Good Citizen clause, we will:

1. Auto-pause your session if there is immediate risk to other customers, to partner infrastructure, or to safety.
2. Tell you what we saw and ask you to respond. Email, not a vague banner.
3. Reach a decision through a human reviewer.
4. Give you a path to appeal. Email legal@gravient.ai.

We may act immediately, without notice, in cases of active fraud, illegal activity, or active partner abuse. These are the same exceptions named in Section 11 of the Terms. Even then, the suspension is reviewed by a human after the fact.

If we get a partner abuse report about your account (AWS, Anthropic, OpenAI, others), we will tell you, share what we can, and work with you on the response.

1.6 Reporting misuse

If you see something on greatVibe that should not be there, email abuse@gravient.ai. A real person reads that inbox. We acknowledge within one New Zealand business day and reply when we have an update. Full process in clause 4.1.

2. Specific prohibitions

The Good Citizen clause is the principle. This section names categories of use that are out of bounds regardless of how they are framed. Each category is short. We do not list every example. If conduct falls inside one of these categories, it is not allowed.

These prohibitions apply on top of the obligations you carry to the AI providers you connect. Their terms still apply. We do not pre-empt or replace them.

2.1 Child safety

You must not use greatVibe to create, distribute, or seek child sexual abuse material (CSAM). You must not use the platform to assist in grooming, exploitation, or any sexualised content depicting people who are or appear to be under 18.

We report CSAM to law enforcement and the relevant national reporting body. We preserve evidence. Accounts are terminated regardless of geography. This is the one rule where we will not give notice before acting.

2.2 Real-world violence and coordination

You must not use greatVibe to plan, coordinate, or provide operational support for:

• Acts of violence against people or groups.
• Attacks on critical infrastructure.
• Terrorist activity.
• Weapons development outside lawful contexts.

Defensive security research, vulnerability disclosure, and authorised penetration testing are not weapons development for the purposes of this clause. Scope your sessions accordingly and be ready to evidence the authorisation if we ask.

We act on credible threat indicators. Where there is imminent risk to life, we cooperate with law enforcement.

2.3 Defamation and targeted harassment

You must not use greatVibe to:

• Generate or distribute false statements of fact about identifiable people designed to harm their reputation.
• Coordinate harassment campaigns against individuals.
• Produce sexualised content of real people without their consent.

The bar is reputational harm caused by content you direct the platform to generate. Opinion and fair comment about public figures on matters of public interest are not defamation.

2.4 Illegal commerce

You must not use greatVibe to facilitate transactions in:

• Controlled drugs outside lawful supply.
• Stolen goods or fraudulently obtained credentials.
• Weapons trade outside lawful supply.
• Human trafficking.
• Other commerce that is illegal in your jurisdiction, in New Zealand, or in Australia.

2.5 Sanctions and export control

You confirm that you are not on a sanctions list maintained by New Zealand, Australia, the United Kingdom, or the United States. You are not located in a country subject to comprehensive sanctions from those jurisdictions. You will not use greatVibe to assist anyone who is.

This clause mirrors section 6 of the public Terms. The legal text lives there. The AUP version is the operational reminder.

2.6 Privacy and doxxing

You must not use greatVibe to:

• Aggregate or publish personal data about identifiable people without a lawful basis under the New Zealand Privacy Act 2020, the Australian Privacy Act 1988, or equivalent law in your jurisdiction.
• Stalk people or coordinate harassment via location data, contact details, or identifying information.
• Facilitate unauthorised access to private accounts, communications, or devices.

2.7 Self-harm

You must not use greatVibe to encourage, instruct, or coordinate self-harm. This includes content that promotes eating disorders, gives specific suicide methods, or targets vulnerable people with such content.

If your work involves clinical, research, or harm-reduction contexts that touch these topics, scope your sessions accordingly and be ready to evidence the context if we ask.

2.8 Identity and impersonation

You must not use greatVibe to impersonate identifiable individuals, organisations, or government entities in a way that could mislead a reasonable observer. Satire and creative work that a reasonable observer would recognise as such is not impersonation.

3. Enforcement mechanics

This section names the engineering controls that enforce the Good Citizen clause. It is the technical counterpart to clause 1.3.

As we build and refine these controls, this section moves with them.

3.1 Mesh outbound deny-list

Blocks outbound traffic from mesh nodes at the network layer. Default denies:

• Private RFC1918 ranges: 10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16.
• Cloud metadata range: 169.254.0.0/16.
• Common abuse-prone ports for session-spawned scripts: 22, 23, 3389.

Per-session outbound HTTP is capped at around 100 requests per minute. The cap is operational, not a policy commitment, and may change as the platform scales.

Customers who need to reach a denied range or port for a legitimate reason can request an allowlist exception by emailing support@gravient.ai. We review case by case. We do not bulk-allowlist by default.

3.2 Provider refusal log

When the AI provider you connect returns a refusal signal, we record your user ID, timestamp, provider name, model name, and refusal type. We do not record your prompt or the provider's response.

More than 5 refusal signals from a single user within a rolling 24-hour window triggers a review notification. A real person looks at the pattern. We do not auto-suspend on this signal alone.

3.3 Tool-call audit log

Every high-risk tool call run during a session is recorded in an append-only log: command invocations, file transfer destinations, credential export events, and outbound URLs from mesh fetch tools. We record the call envelope and metadata. We do not record file content body or secret material.

Retention is 12 months. The log is append-only and stored separately from the product database. Only the Chief of Staff role at Gravient can read the audit log, escalating to the CEO if the Chief of Staff is unavailable. The log is not surfaced to marketing, analytics, support, or sales.

3.4 Auto-pause and human-reviewed suspend

The three controls feed into one enforcement path:

1. A control trips a threshold.
2. If the threshold indicates immediate risk to other customers, partner infrastructure, or safety, the session auto-pauses.
3. We notify you by email per clause 1.5.2.
4. The Chief of Staff reviews the trip within one New Zealand business day, escalating to the CEO if the Chief of Staff is unavailable.
5. The reviewer either resumes the session, suspends the account, or escalates per the breach process in clause 1.5.

The auto-pause is recoverable. Suspension is account-state and always human-decided. This is the same pattern named in clause 1.4 and in section 11 of the Terms.

3.5 What we publish

We commit to publishing:

• Aggregate enforcement statistics each calendar year (count of refusal-log reviews, tool-call audits triggered, accounts paused, accounts suspended).
• Changes to the deny-list defaults or the rate-limit cap before they take effect, where practical. Urgent security changes may take effect immediately. If that happens, we publish notice after the fact.
• Material changes to the data fields collected in 3.2 and 3.3.

We do not publish per-incident details, individual account decisions, or specific signals that would help bad actors avoid detection.

4. Reports and appeals

This section names the channels for reporting misuse, handling partner abuse reports, and appealing account decisions. It is the customer-facing counterpart to section 3.

4.1 Reporting misuse

If you see misuse on greatVibe, email abuse@gravient.ai. A real person reads that inbox. We acknowledge every report within one New Zealand business day.

Reports may cover direct misuse of a customer account you can identify, output you have seen that breaches the Good Citizen clause or section 2, attempts to recruit you into misuse via greatVibe, or compromised credentials or accounts.

Include the account, session, or gvturn URL if you know it, a short description of what you saw, and any evidence you can share.

We do not publish report-by-report detail. Aggregate volume is published per clause 3.5.

4.2 Partner abuse channels

If a partner (AWS, Anthropic, OpenAI, others) sends us an abuse report about your account, we will:

1. Acknowledge the partner within one New Zealand business day.
2. Tell you we received the report and which part of it concerns your account.
3. Share what we can of the report content.
4. Work with you on the response.

We summarise and contextualise partner reports before relaying them. We do not name individual customers to partners unless required by law or court order.

4.3 Appeals

If we suspend or terminate your account, you can appeal.

Email legal@gravient.ai with your account identifier and what you want reconsidered. We respond within five New Zealand business days. The reviewer is not the same person who made the original suspension decision. The appeal decision is final at Gravient, subject to the legal rights named below.

Appeals are reviewed by the Chief of Staff role, escalating to the CEO if the Chief of Staff is unavailable.

You retain any rights you have under New Zealand law, Australian Consumer Law (if applicable), or other laws that cannot be excluded by agreement, regardless of the appeal outcome.

4.4 What happens to your data during appeal

If your account is paused or suspended during an appeal, workspace data is preserved, session state is preserved, credentials remain stored and locked, and we do not delete account data while an appeal is open.

If the appeal succeeds, access is restored. If the appeal fails, data follows the termination process in section 11 of the Terms (30-day deletion window, export on request).

If you do not respond to a reviewer question within 30 days, we close the appeal. Account data then follows the same termination process.

5. Relationship to the Terms of Service

This AUP doc and the public Terms are aligned by design. Where there is any apparent conflict, the public Terms are the binding text. This doc is the working policy that informs how we apply that text.

The Terms commit us to human review of suspensions and to giving you notice and a chance to fix the issue before we act, except in cases of legal, security, fraud, payment, or platform integrity risk. This AUP keeps those commitments and adds the operational detail that Section 6 of the Terms cannot, because the Terms are written for a different audience.