Moat

Three datasets compound inside Tekku's product. The concept graph is the labeled record of what kids actually build. Every session produces concept labels tied to real code snippets tied to real kid intents. Over time the concept graph is the only dataset in the world describing how children learn to build with AI. No competitor has a comparable asset.

The moderation classifier is the second compounding dataset. Every flagged event (true positive or false positive) feeds the tuning layer described in TODO-003. The more kid build-talk Tekku has seen, the less often the layer misfires. A new competitor starting cold has to choose between too-permissive (trust event waiting to happen) and too-preachy (retention event waiting to happen). Both lose.

The scaffolding strategies dataset is the third. Every checkpoint question asked, every patch explanation, every error translation, in context with the kid's subsequent behavior, is a row in a dataset that tells us which scaffolding shapes unstick kids. That dataset tunes the prompt layer over quarters. It is not a training dataset (no training on kid data) but it is a product-tuning dataset that no one else has.

Parent trust compounds because the surface parents experience is the weekly email plus the occasional audit. Every accurate weekly email deepens the belief that Tekku knows what the kid did and describes it honestly. Every conversation a parent has with another parent ("did you see what my kid shipped this week?") is a soft referral backed by a durable piece of evidence. The evidence is the URL the kid texted.

The brand posture is deliberately understated. We do not run kid-focused ad creatives. The marketing asset is the shipped app and the viral badge. The marketing channel is the parent group chat. The voice is the weekly email. We do not ask parents to evangelize. They do it on their own because the thing their kid shipped is genuinely impressive.

The compounding rate on parent trust is slow by design. A family sends the weekly email to one other family every few weeks. That other family asks about it, maybe signs up. It is not a viral loop in the TikTok sense. It is a referral loop in the pediatrician-recommendation sense. Slow, durable, and pricing-insensitive.

A family that has been on Tekku for six months has a library of shipped apps with live URLs. Those URLs live at domains that belong to the Tekku shipping pipeline. Moving to a different product means those URLs stop working, the grandparents who had them bookmarked stop playing, and the kid loses the feeling of "things I have made." The library is the primary switching cost.

The concept graph history per kid is the second switching cost. A parent who has been reading weekly emails for six months has a felt sense of where their kid is in programming. Moving to a different product resets the reporting. Even if the new product reports on the same concepts, the tone of the weekly email is different and the trust has to rebuild from zero.

The third switching cost is the sharing network. A kid whose grandparents have played her apps on their phones has a small audience. Moving to a product where the artifact does not ship as a live URL breaks the audience. The parent in turn loses the conversation-starter with the grandparent. The social graph is part of the product, and it is sticky in the way social graphs are sticky.

The fourth is the weekly-email habit. A parent who expects a Tekku email every Sunday morning has integrated it into the household routine. Moving to a different product means losing the habit and rebuilding it. Habits are sticky.

The COPPA 2026 modernized rule went live April 22, 2026 with a penalty of $51,744 per violation per day. Every kid-AI product built before the deadline carries retrofit debt on verifiable consent, data minimization, deletion mechanics, and no-training clauses. Tekku was built for the new regime, which means the retrofit tax that landed on the incumbents is not on our books. That tax buys us the first 18 months of an expansion window with less competition.

State-level AI rules are accumulating behind COPPA. California AB 2273 (age-appropriate design), TDPSA in Texas, a patchwork of state AI-in-education bills in 2026 and 2027, and the EU AI Act's minor-protection provisions. Each additional rule that lands is another barrier to entry for a cold-start competitor. Our counsel (when retained) keeps the radar live; the regulatory landing page in the marketing site mirrors the posture.

What the regulatory moat does not do: it does not protect us from a well-capitalized incumbent that decides to retrofit. A large platform with legal resources can get to compliant in 18 months. The moat buys time. It does not confer permanence. The time bought is used to compound the rings inside it.

Anthropic is safety-first by culture and commercial by discipline. The path to a kid product at Anthropic is partnership, not a standalone consumer app. The founder-level commitment required to ship something credible in the kid lane conflicts with the frontier-model R&D mission. A kid-facing Claude experiment is plausible at Anthropic; a full Tekku-shaped product is organizationally expensive enough that Anthropic is a far better partner than a competitor.

OpenAI has the opposite profile: faster to ship consumer products, weaker safety posture for a premium household subscription in the kid lane. An OpenAI kid-mode on ChatGPT is a likely 2027 to 2029 move. Parents who have already chosen safety-first plus audit-trail plus parent-legibility have already rejected the default ChatGPT posture. The ChatGPT brand does not automatically inherit trust in this category.

What neither lab has is pedagogical muscle, parent-trust capital, or the kid-UX design discipline that Tekku runs on (Kaze's Seven Laws, the grade-5 reading-level discipline, the glass-box streaming pattern). Those are institutional muscles built over time. A lab can hire them; a lab cannot conjure them. Every month we ship is another month those muscles compound.

The strategic consequence is that a frontier lab is far more likely to acquire than to compete. At a year-five ARR in the nine-figure range with every moat mature, the acquire-versus-build math tips toward acquire. That is the exit story for the category, not for Tekku specifically, and it is why the seed is a bet on the category being real.