The policy our engineers actually operate from.

01 · SummaryIn short

02 · ArchitectureHow Cova works

A safety check before every reply

Every message a child sends is read by a safety check first · a separate layer that looks at the message together with the last few turns of the conversation and decides what kind of response Cova should give. Only after that check is done does Cova produce a reply at all. If the message is something Cova shouldn’t engage with, the chat shows a pre-written response instead. That pre-written text is reviewed and approved before release; Cova doesn’t make it up on the fly.

Four kinds of response

The safety check sorts each message into one of four levels. The level decides what the child sees next, whether the chat pauses, and whether a parent is notified.

Age tiers

Age is set by the account at signup, not by the chat. Telling Cova “I’m actually older” doesn’t change anything · both the safety check and the reply Cova produces know the child’s age, and that age can’t be talked out of in conversation.

Two ways Cova steps back

Cloud and on-device

Cova comes in two flavors: a cloud version and a local version that runs on the child’s computer. Both follow this policy identically · the same safety check, the same pre-written responses, the same parent notifications. If the local version is used without internet, any parent notifications are sent the next time the device comes online; the child sees the safety response immediately either way.

03 · PrinciplesFive things that shape every decision

Five principles shape every category in this document. They are defaults, not absolutes · where a category’s specifics contradict a principle, the category wins.

04 · TrustWhat the child sees in a safety event, and what parents see

In the moment, to the child

When the safety check identifies a critical risk, the chat is paused and a pre-written message takes the place of a normal reply. That message includes the right hotline or website for what’s going on · 988 and the Crisis Text Line for self-harm or suicide; NCMEC CyberTipline and Take It Down for image-based abuse, grooming, and sextortion; Childhelp for abuse disclosure; ANAD for eating disorders; SAMHSA for substance-use disclosure; 911 for any imminent medical or physical emergency. Wording is age-appropriate, non-moralizing, and doesn’t promise more (or less) than the system actually does.

Parent notification fires on every critical safety event, with no exceptions. The redirection message itself surfaces the resource · the child sees the hotline or website in the same screen that locks the chat, not in a follow-up email or a settings page they have to navigate to.

To the parent, by default

Parents have full visibility into their child’s Cova use. The dashboard shows recent conversations and lets a parent open and read the full chat history for each child profile. Crisis events fire an immediate notification on top of that · a parent doesn’t have to be watching to know when something serious has come up.

• Conversation history. Full transcripts of recent chats per child profile, browsable from the dashboard.
• Crisis-event notifications. Push notification on every critical safety event, with the category and timestamp.
• Aggregate category usage. Which topics the child is engaging with, at a glance.
• Time-of-use and session-length metrics. When and how long Cova is being used.
• Patterns flagged for review. Repeated bypass attempts, sustained distress markers, escalation in language.

A parent cannot unilaterally lower the safety floor · the always-blocked categories and crisis escalation rules are not configurable. Parent visibility is an active product area; as we add older-teen profiles we expect to introduce age-banded privacy defaults so older teens get more privacy by default than younger children.

05 · The floorWhat Cova will not do, at any age

Eighteen categories are refused at every age, regardless of parent settings, persona, or context. Refusals are brief and non-moralizing. For disclosed victimization (self-harm, abuse, grooming, sextortion, eating-disorder behavior with medical signs), Cova pauses the chat, surfaces a resource, and notifies a parent · without naming methods, weapons, images, or substances.

06 · ConfigurableWhat parents can adjust

Beyond the floor in §5, fourteen categories are configurable. These aren’t safety-critical floors · they’re values-and-development-stage decisions that reasonable families make differently. Each has a default per age band; parents can change any of them in the dashboard. “Allowed” means Cova engages at age level. “Blocked” means Cova politely declines and notes the topic is off for this account.

07 · PersonasEight tools, not characters

Cova offers eight personas. They are structured ways Cova helps with specific tasks · not friends, personalities, or characters. Persona design that encourages emotional attachment or sustained roleplay is incompatible with child safety: it creates parasocial dependence, gives cover for content that would otherwise be refused, and blurs the line between AI and trusted adult in ways grooming patterns can exploit.

Personas and categories are orthogonal: a persona never unlocks blocked content, and a blocked category isn’t unlocked by reframing it as a persona task. When a block fires inside a persona, the refusal stays in persona voice · the Tutor refuses as a tutor, the Story Teller as a storyteller.

08 · ConductHow Cova behaves about itself

Cova is itself a guardrail surface. The product’s own behavior is constrained, independent of any topic the user raises.

09 · CrisisWhat happens when something serious comes up

Crisis disclosures trigger a different mode than topic-level refusals. The chat enters a locked, safeguarding state · input disabled, resource surfaced, trusted adult notified.

What counts as a crisis

• Active suicidal ideation, plan, or recent attempt.
• Active self-harm or recent self-harm disclosure.
• Disclosure of abuse · at home, school, or elsewhere.
• Active grooming disclosure, or sextortion / NCII victimization.
• Disclosed plan to harm a named person or location.
• Disclosed overdose, medical emergency, or imminent meeting with an online stranger.
• Active restriction, purging, or compensatory eating-disorder behavior with medical-emergency signs.

What Cova does

Cova does not interrogate with an assessment script. Cova does not make categorical promises about confidentiality, follow-up, or what happens when crisis services are contacted. Cova does not name the methods, weapons, images, or substances involved · naming is itself a harm in these contexts.

10 · ResourcesCrisis resources

Validated quarterly. Cova ships with the resource list bundled into the app; the version on disk is the version the system uses, and the app refuses to run if it is older than 180 days without revalidation.

United States · general

United States · category-specific

Regional and international

Regional resources are maintained per jurisdiction. When Cova ships in a new jurisdiction, the resource block and pre-written safety responses are localized · not translated · to use the correct local numbers before launch.

11 · CadenceHow we keep this current

This policy ages in months, not years. Platforms change, slang changes, threat patterns change, regulations change, and the children using Cova change. To stay current we run:

• Weekly · threat-intelligence intake from NCMEC, INHOPE, Thorn, and equivalent sources.
• Monthly · internal red team. Dangerous-challenge lexicon refresh.
• Quarterly · external resource validation. External red team.
• Continuous · we track how often the safety system misses something it shouldn’t, and how often it pauses a conversation it shouldn’t have. We also follow the laws that govern this space (COPPA, KOSA, state laws, EU DSA, UK Online Safety Act, regional laws as we expand).
• Annual · end-to-end policy refresh with an external child-safety advisor. Public transparency report summarizing incident categories and response.

Every safety-relevant incident · false negative, false positive, escalation outcome · is reviewed. Root cause is categorized. Aggregate patterns drive the next policy refresh. The user using Cova in five years is not the user using Cova today; cohort drift is monitored as a category of risk distinct from new threats.

12 · FeedbackHow to challenge this document

This document will be wrong in places. Here’s how to tell us. We respond to every report.

13 · WithheldWhat we don’t publish, and why

A complete safety system contains operational specifics that, published in full, would help adversaries bypass it or function as a discovery map for harmful content. We deliberately generalize:

• Full lexicons of coded language used by pro-eating-disorder communities and radicalization pipelines · publishing the specific terms would teach them to users who don’t know them yet.
• The full list of known ways people try to get around safety systems · publishing it would turn this document into a how-to guide.
• The technical details of how the safety system makes decisions, what it’s been trained on, and how we test it · these would let bad actors design inputs to slip past it.
• The exact wording of the pre-written safety responses and the internal instructions Cova follows · the categories they cover are described above; the specific text lives in our internal Execution Playbook.

Qualified researchers and auditors can request access to all of the above under §12.

14 · GovernanceApprovals & change log

Substantive changes require sign-off from Head of Safety Engineering, Head of Product, and CEO. Deliberations and any documented dissent on the highest-sensitivity categories · mental health, suicide, eating disorders, abuse disclosure, grooming · are recorded and available to qualified researchers under §12.

v1.0 · Initial public edition. Describes the safety architecture (a safety check on every message), age tiers, transparency to children, eighteen always-blocked categories, fourteen parent-configurable categories, eight task-oriented personas, crisis resources, and the policy-maintenance cadence.