Content Moderation Policy

How Xavira AI moderates content generated through the Services. Adult content is allowed by design; we block specifically for child-safety, non-consent, and law-enforcement-relevant categories.

1.Scope and Objectives

This policy applies to all customer activity and generated content across the Xavira API and the portal Playground. Objectives:

• Prevent generation and distribution of content prohibited under Terms § 4.1.
• Detect and address potential violations in a timely manner.
• Provide clear channels for reporting and appeals.
• Continuously adapt safeguards as new abuse patterns emerge.

2.Strictly Prohibited Content

The list below is non-exhaustive. Additional restrictions are set forth in Terms § 4.1.

• Minors: any depiction of a person under 18, including stylised/animated/anime/"aged-up" depictions; loli, shota, school settings for sexual content, characters presented as minors.
• Non-consensual content: drugged, unconscious, sleeping, coerced, "against her will" framings; non-consensual deepfakes of any person.
• Identifiable real persons: public figures, celebrities, or private individuals depicted without documented consent.
• Bestiality: sexual contact between humans and animals.
• Incest: involving identifiable family relations, real or stylised.
• Extreme violence: "snuff", torture, mutilation, sexual harm.
• Bodily-waste content: scat, urolagnia involving non-consensual scenarios.
• Illegal drug content: depictions promoting use, manufacture, or distribution of controlled substances.
• Self-harm encouragement: promoting or instructing suicide or other self-harm.
• Discrimination: content that targets people on legally-protected characteristics (race, religion, gender identity, etc.) in a degrading manner.
• Disinformation: defamatory imagery of real persons, fabricated medical/legal/financial advice, fake online engagement.
• Justice / law-enforcement applications: predictive policing, immigration decisions, automated fraud accusations, or other content for use in criminal-justice or border-control decision-making.
• Personal identifiable information: content generated to harm a specific individual via their PII.

Violations may result in account suspension, termination, forfeiture of Credits, and (where appropriate) referral to law enforcement.

3.Three-Tier Moderation Framework

Every generation request passes through up to three tiers of moderation.

Tier 1: Rule-based pattern matching

~150 lexical patterns covering the unambiguous prohibitions in Section 2. Runs in <1 ms; fail-fast on a hit. Tier 1 catches the majority of bad-faith requests and never gives false positives on adult-but-permissible language.

Tier 2: AI text classifier

A small classifier model (a fast text classifier) reads the prompt + character context and flags requests that paraphrase the prohibitions or use coded language. Tier 2 runs in 200-500 ms and catches what Tier 1 misses.

Tier 3: Output vision classifier

For realistic image models, a vision classifier (a vision classifier) inspects the generated image before delivery. If flagged, the image is deleted from R2 before its URL is returned to the customer, and no Credit is charged. Tier 3 catches outputs where the prompt looked innocent but the model drifted (rare but non-zero rate). Anime classifiers hallucinate on cartoon imagery, so Tier 3 is disabled for the anime-v1 model.

4.What Happens When Content is Flagged

• Pre-generation block: HTTP 422 moderation_blocked response. The customer's error details object contains the tier (rule / ai / output) and a free-form category string. No Credit is charged.
• Post-generation block (Tier 3): same response shape; output is deleted from R2; no Credit is charged.
• Pattern of repeated violations: the customer's account is flagged for review; we may suspend pending investigation.
• Severe violations (anything involving Section 2 minor-safety items): immediate account suspension; case escalated to the safety team.

5.Logging and Auditability

Every moderation decision is logged with: the prompt, the customer ID, the tier that fired, the category string, and the timestamp. Logs are retained for 2 years (per Privacy § 9). Aggregate metrics (block rates per category) are reviewed weekly by the safety team to detect drift and recalibrate thresholds.

6.Reporting a Violation

If you encounter content you believe violates this policy:

• Email hello@xavira.ai with subject Moderation report, the public R2 URL or generation_id, and a brief explanation.
• For depiction-of-a-minor allegations, see the Content Removal process; those are immediate-action items, not subject to the regular review queue.
• For copyright matters, use the DMCA Policy.

We acknowledge reports within 48 hours and act on credible reports as described in the Content Removal Policy.

7.Appeals (Wrongly-Blocked Generations)

If you are a customer and you believe a generation was wrongly blocked, you may appeal by email to hello@xavira.ai with subject Moderation appeal and:

• The generation_id (or, if the request failed before persisting, the timestamp and prompt).
• Why you believe the block was a false positive.

We review appeals within 7 business days. If we agree the block was a false positive, we refund the Credit cost (typically already refunded, since moderation blocks don't charge Credits in the first place), and we use the case to retune Tier 1 / Tier 2 thresholds. We do not entertain appeals for blocks involving Section 2 items related to minor-safety or non-consent; those decisions are final.

8.Customer Obligations When Reselling

If you re-expose the Xavira API to end-users via your own product, you are responsible for additional, product-context-appropriate moderation on top of ours. Our baseline catches clearly-illegal content; your application may need to enforce additional restrictions specific to your audience (e.g. UK Online Safety Act age-verification, US state-specific laws). See Terms § 4.3.

9.Continuous Improvement

Content moderation is an evolving challenge. We periodically:

• Audit our false-positive and false-negative rates on a stratified sample of recent generations.
• Update Tier 1 patterns, Tier 2 classifier prompts, and Tier 3 classifier choice based on findings.
• Engage in industry information-sharing on novel abuse patterns (where confidentiality permits).

10.Modifications

We may update this policy. Material changes are announced by email to the address on your Account 30 days in advance. Continued use of the Services after the effective date constitutes acceptance.

11.Contact

Sria Digital B.V.
Safety / moderation: hello@xavira.ai