Risks

Stage 1 ships with user number one as the founder's own son, so verifiable parental consent is informally satisfied. The moment user number two joins (Week 3 Day 17 per the Stage 1 master plan), verifiable parental consent, 90-day retention, parent-initiated deletion, and a no-training clause all become hard legal prerequisites. A modernized COPPA rule lands April 22, 2026, and penalties reach $51,744 per violation per day.

Mitigation: TODO-001 in the project backlog is the implementation ticket. Scope: Persona or Stripe Identity integration, privacy policy drafting, Supabase retention cron, parent-facing export and delete UI. Estimated one week of focused work before Stage 2 onboarding.

Gate Stage 2 family onboarding on TODO-001 merge. No non-founder kid in the system until consent infrastructure is live.

Outside counsel retained for quarterly privacy audit once infrastructure ships.

Safe-harbor filing with an FTC-approved program targeted at Stage 2 exit.

The EU AI Act carries elevated obligations for AI systems that interact with minors, including transparency, human-in-the-loop requirements, and restrictions on certain categories of inference. The Act is not a current blocker because Tekku does not serve EU kids in Stage 1 through Stage 3. It becomes a gating item at the international expansion vector.

Mitigation: Mitigation: Do not onboard EU kids until a localized compliance package is ready. Geo-gate signup at the IP and payment-method level.

When EU is in scope, retain EU counsel and complete a conformity assessment before launch.

Architecture choices made today (no training on kid data, data minimization, cascading delete) are compatible with the Act; the gap is paperwork, not design.

Multiple US states are drafting or advancing legislation that governs AI use in K-12 classrooms. Bill text varies by state and is actively changing. Because Tekku's Workshop tier enters schools in Stage 2, state-level requirements become a pilot-deployment gate in every district we contract with.

Mitigation: Mitigation: Track active bills in the states where a Workshop pilot is contracted. Do not sign a pilot in a state whose pending legislation we cannot comply with at launch.

Every district contract includes an AI-use rider describing model calls, data flow, moderation, and retention. The rider is designed to satisfy the strictest current state rule, so adding a state is a contract check, not a rebuild.

TODO(counsel): confirm current bill status per target-state list before Stage 2 pilot signatures.

Workshop tier activates FERPA. Tekku operates as a school official under a written contract with each district, with defined educational purposes and limited use of educational records. The edge cases (shared classroom devices, incidental capture of non-enrolled kids, transcript portability between districts) are where FERPA exposure concentrates.

Mitigation: Mitigation: District contracts explicitly name Tekku as a school official and scope the educational-records use. No record leaves the district namespace.

Moderation transcripts are logically separated per district; no cross-district aggregation for any purpose.

Parent consent for Workshop is collected by the school under state law. Tekku maintains the technical controls.

TODO(counsel): template district contract reviewed by outside counsel before Stage 2 pilot.

Claude is the primary model for kid-facing chat and the submit_patch tool. An Anthropic outage, a rate-limit event at a peak hour, or an unexpected policy change on the consumer product would degrade or halt the kid loop. No secondary provider is wired in Stage 1.

Mitigation: Mitigation: Server-side circuit breaker degrades to a static "Tekku is resting, try again in a few minutes" flow rather than a hard error. Parents see a status badge. The degradation is a controlled product state, not a crash.

Enterprise contract with Anthropic targeted before Stage 2 exit to secure committed capacity and a no-training tier.

Architecture keeps the Claude call isolated behind a single adapter, so a secondary model (OpenAI, a future DeepSeek apprenticeship) can be swapped in at the adapter layer without touching the product.

Observability: Sentry alert on any 4xx/5xx rate above a threshold for the Anthropic call, paged to the on-call engineer.

Claude can confidently produce wrong code, wrong explanations, or wrong encouragement. For a kid-facing product, a single confident-but-wrong explanation is a trust event with the parent, not just a bug. The risk is structural to the technology, not specific to one provider.

Mitigation: Mitigation: The submit_patch tool contract forces Claude to return structured patches, not free-form code. Patches run through a Babel AST validator that whitelists known imports before they reach the kid's sandbox.

Stage 1 uses full-file replacement with server-side validation of the template whitelist. A hallucinated import is rejected before render.

Parent-facing weekly email shows what the kid built, not what Claude said. The claim the parent sees is a ship event, a verifiable artifact.

Ongoing prompt review: the cached system prompt is versioned and changes ship with dated diffs for traceability.

A kid sends a prompt, moderation passes it, Claude generates a response that contains content not suitable for the kid. A single confirmed false-negative is a high-severity trust event and a potential enforcement trigger. Probability is low because two moderation passes run (pre- and post-generation), but the impact is real.

Mitigation: Mitigation: Every kid message passes through OpenAI Moderation before it reaches Claude. Every Claude response passes through the same filter before it reaches the kid.

Content policy in the cached system prompt constrains Claude to build-appropriate output, which reduces the load on the post-moderation filter.

Flagged events are logged to a moderation event table with full text snippets and reviewed inside a 24-hour human queue for the first six months.

Incident playbook: parent notified within one hour of confirmation, account frozen, root cause documented on the public transparency page within the week.

OpenAI Moderation flags kid build-talk as violent. "Kill the timer." "Shoot the ball." "Hit the enemy." Without tuning, kids feel the tool is preachy or broken. A single viral screenshot of a legitimate prompt being refused can slow signup for weeks. This is the risk that showed up in concept testing.

Mitigation: Mitigation: TODO-003 in the project backlog is the tuning-layer ticket. Scope: secondary classifier over the moderation API, allowlist of verified-benign phrases with 90-day expiry, public transparency page.

Stage 1 logs every flagged event to the moderation_events table. After four weeks of data, phrases flagged three times with no actual harm pattern enter the tuning queue.

Human review queue for ambiguous events during the first six months. Ambiguity becomes a classifier example, not a kid-blocking decision.

Public /safety transparency page documents what the current policy allows and why. The second false-positive is a process failure, not a discovery.

A frontier-lab consumer kid product would compress the category timeline. The probability is lower than most investors assume because kid verticals carry compliance overhead that frontier labs historically decline to absorb, but the impact case is real. The risk window is the two years before Tekku's category ownership is established.

Mitigation: Mitigation: Compliance-first architecture is a specific kind of insulation. A generic kid-chat product shipped by a frontier lab without a COPPA-era plan runs into enforcement before Tekku does.

Brand and pedagogy: Tekku is not a chat wrapper, it is a build loop with a Teko mascot, a gamification layer (see the Tekku gamification plan), and a parent weekly review. That is hard to replicate with a horizontal product.

Build the acquisition case: a kid-education brand with FTC-safe harbor, proven retention, and signed school pilots is the category asset a frontier lab would rather buy than build.

Khan Academy's Khanmigo is the most credible incumbent AI tutor. If Khan adds a kid-building or project-creation surface to its product, the Workshop school-sales motion runs into a trusted name with existing district relationships.

Mitigation: Mitigation: Direct-to-parent is the primary motion in Stage 1 through Stage 3. Workshop revenue is not the foundation; it is an ARR accelerator.

Tekku ships what kids build. Khan ships what kids learn. The output artifact, a working app a kid can send a grandparent, is a structurally different product experience.

Speed: Tekku's Stage 1 through Stage 3 cadence ships parent-facing features weekly. Khan's product cycle is slower by design.

The Workshop tier is structured so that a district already using Khanmigo can add Tekku in the creative-building slot without displacing the tutor.

A major press event involving a kid-AI product (harm incident at a competitor, a regulatory investigation, a documentary) can freeze parent willingness to pay for any kid AI product for a quarter or two. This has happened in adjacent categories, including social apps.

Mitigation: Mitigation: Compliance-first and transparency-first positioning is defensible in a backlash. Parents choosing among kid-AI products in a cautious moment pick the one with an FTC safe-harbor filing and a public transparency page.

The subscription model means runway is not dependent on a growth rate in any single month. A freeze on paid acquisition for a quarter extends the runway consumption timeline.

The weekly parent email, delivered whether or not the category is in the news, keeps existing parents anchored to the specific value their kid is getting.

Risk case financials still survive on runway alone (see scenario analysis on the financials page).

Anthropic raises prices, the cache hit rate drops, or a bug causes a session to loop. A single bad day could spike spend well above the model assumption. Permanent pricing changes would compress gross margin at scale.

Mitigation: Mitigation: Hard $200 per day platform cap and $30 per day per-parent cap in the Claude invocation layer. Caps trip before a real financial event lands, with a PagerDuty alert at 80%.

Per-session cost cap of $3.00 (with $1.50 to $3.00 expected spend) prevents a looped session from running away.

Enterprise contract negotiation targeted at Stage 2 exit to secure committed pricing.

If Anthropic raises prices permanently, the subscription price tier ($149 Stage 1, A/B against $199 in Stage 3) carries the gross margin with pricing power.

One founder today. Founder illness, burnout, or departure halts the company. Institutional due diligence will surface this. Seed conversations are harder without a plan.

Mitigation: Mitigation: Seed capital funds two founding engineering hires plus a head of trust within the first 120 days post-close.

Key-person insurance on the founder taken out at seed close.

Documentation posture: the BuildOS agent team holds operating context, plan documents live in docs/plans, and the system prompt and tool contracts are versioned. The bus-factor exposure compounds down, not up.

Quarterly sabbatical policy announced at seed close: one week offline per quarter, no exceptions. Burnout as a failure mode is designed against, not hoped against.

Parents are a deliberative audience that asks specific questions. At 1,000 families the support volume is manageable by the founder. At 10,000 families it is not. Support overhead scaling faster than revenue is a realistic operational drag on gross margin.

Mitigation: Mitigation: Staged hire. Head of trust role is in the seed hiring plan at month four post-close, before the Stage 3 growth curve steepens.

Parent-facing self-serve: export data, delete account, change tier, pause subscription. Every high-volume ticket type has a product solution before it becomes a ticket type.

Weekly parent email includes a one-click "reply with questions" flow that routes to a triage inbox with tagged categories. The inbox becomes the product-feedback loop, not just a cost center.

SLA: two business-day response target at Stage 1 through Stage 2, 24-hour at Stage 3+.

The base case assumes a specific retention curve. If monthly churn runs at 12% instead of a target near 9%, the model still works but the payback period extends materially. If churn runs to 15% the model breaks. The parent demographic is retention-sensitive to weekly proof of value.

Mitigation: Mitigation: The weekly parent email, the twelve workspace states, the badge ship events, and concept-based progression are all retention multipliers. If the curve slips, reinvest in the email and ship flow before touching pricing.

Founding-100 cohort proves retention on a high-quality known group before paid acquisition activates.

Product telemetry on retention is segmented by signup cohort, so a pricing-test effect is distinguishable from a product-change effect.

Runway sensitivity is modeled on the financials page. The conservative case is still a defensible venture outcome at 12% monthly churn.

Paid acquisition cost for parents of kids in the target age range is not a stable number. Facebook and Google paid CPMs change quarterly. If CAC runs 30% above the model assumption, payback period extends and the growth rate tolerable at a given burn compresses.

Mitigation: Mitigation: Organic and referral channels are the primary paid-acquisition insurance. The viral badge ship events, the weekly email forwards, and the public gallery are designed to move the blended CAC down as the user base scales.

Paid spend is gated on a CAC-to-LTV ratio threshold. Spend expands only when the ratio is inside band.

Workshop channel (schools) is a low-CAC acquisition path at the bulk level; it does not rely on paid media.

Pricing power ($149 to $199 A/B in Stage 3) is the second lever: a higher ARPU absorbs a higher CAC at the same payback-period target.

Kid enrichment is a discretionary line item in family budgets. A sustained macro downturn (recession, a persistent cost-of-living shock) compresses the addressable audience. International expansion adds FX exposure on revenue denominated in local currency versus costs denominated in dollars.

Mitigation: Mitigation: Stage 1 through Stage 3 are US-only. FX is not a present concern; it is a Stage 4 gating item.

Pricing tiers include a Family Plan at $249 Stage 3 and a grandfathered $99 for founding families. The price ladder gives demand elasticity headroom in a downturn.

School channel is counter-cyclical: district budgets for education technology are less volatile than family discretionary spend.

At international launch, price in local currency and hedge large revenue-vs-cost mismatches.