Rebuilding education after content scarcity — a strategy and experiment system for the moment when explanation becomes free but learning does not.
Education is not one product. It is a politically protected bundle of seven jobs — custody, motivation, instruction, socialization, assessment, credentials, and social infrastructure. AI attacks the third job immediately and parts of the fifth. It barely touches the rest without new institutions around it.
The thesis: education will be unbundled at the instructional layer long before the custody, legitimacy, and social-infrastructure layers. The cheap thing is now explanation. The valuable thing is arranging a human life so that learning actually happens, proving what happened, and making the proof matter.
The strategy: do not begin by replacing schools, teachers, or curriculum. Run two commercial wedges (ProofWorks Studio, Feral Learning Lab), one bounded regulated discovery track (IEP Shadow Ledger), and one internal primitive (Mastery Spine) — each with paid gates and kill criteria.
The whole strategy falls out of one table: what the buyer is really purchasing, and how exposed each job is through 2035.
| Job | What's really being bought | AI exposure | Best attack |
|---|---|---|---|
| Custody & safety | A lawful place for minors while adults work | Low | Give existing custodians better programs and operations |
| Motivation & behavior | Attention, persistence, norms, co-regulation | Medium | Human facilitator copilot, small groups, better feedback loops |
| Instruction & feedback | Explanation, examples, practice, diagnosis | High | Constrained tutors tied to verified materials and mastery models |
| Socialization & belonging | Peers, identity, teamwork, status | Low–medium | Design stronger cohorts and projects; do not simulate friendship |
| Assessment & evidence | A trusted claim about capability | Medium–high | Authentic artifacts, oral defense, longitudinal evidence |
| Credentials & sorting | A socially recognized ticket to the next gate | Medium | Attach verified evidence to existing credentials first |
| Social infrastructure | Special ed, meals, transport, mandated reporting | Low | Automate documentation and coordination around human delivery |
The business implication is sharp: do not sell infinite content into a content glut. Sell reliable learning operations and trustworthy evidence. And the autonomy test is inverted from the demo instinct: an autonomous tutor that completes work for a learner is anti-educational — the metric is what learners can do later, alone, in a changed context.
Let fewer or less-experienced people deliver expert-quality work. Tutor CoPilot's RCT pattern: the AI helped lower-rated tutors most — expert practice compiled into real-time human support, not "AI alone."
The phlebotomy move, ported from the Automation Moats panel: a structured activity with known materials, bounded goals, instrumented steps, and a facilitator. A physical mastery pod is a learning workcell — valid only if the AI can't just hand over the answer.
Machine-readable competencies and rosters are education's accessible service chases. The standards already exist (CASE, QTI, LTI, OneRoster, xAPI, Open Badges) — make them usable, don't invent another protocol.
Schools keep the child; a specialist supplies an unusually good instructional block. Afterschool, homeschool co-ops, camps, and continuing ed are softer entries than district replacement.
A technically superior mastery record means nothing until a consequential gatekeeper accepts it. Sequence: enrich → verified evidence → attach to incumbent credentials → win employer acceptance → only then go independent.
The anchors are enormous and mostly not addressable — which is itself the strategic information.
"A startup cannot multiply students by a list price and call the result TAM with a straight face." 87% of K-12 revenue is state/local, so national policy enthusiasm creates no national buyer — and enrollment is projected lower in 2030 than 2019, so shrinking systems buy workload relief, not experiments.
The founding memo proposed twelve K-8 kits, 6,000 units by Day 30, a marketplace, an outcome guarantee, and a $3.2M pre-seed. The audit's verdict: "a strong strategic intuition wrapped in fictional operating certainty."
| Original claim | What the audit found |
|---|---|
| "$799 bundle, 67% gross margin" | The twelve list prices sum to $1,210 — a hidden 34% discount; margin math holds only at bottom-of-range COGS and before freight, fees, returns, and support. First-order contribution after the claimed $41 CAC: negative $20. |
| "55% Year-2 net margin, $9.4M profit" | Internally inconsistent before ordinary operating costs — the claimed CAC alone consumes $3.3M of the $2.1M remainder. |
| "1.5 years of growth guaranteed" | Undefined across subjects, tests, dosage, and populations; fails FTC ad-substantiation requirements. Insurance does not manufacture efficacy. |
| "Angle Arena laser swords" | Child-operated lasers rejected outright — FDA Class-I guidance for children's laser toys; strings and printed lines teach the same construct. |
| "Anonymized mastery vectors as data moat" | Child learning histories stay linkable and sensitive; COPPA (amended 2025), FERPA, and state law make surveillance exhaust a liability, not a moat. |
| "Patent the pod handshake" | Education interoperability standards already exist; a closed protocol undermines the adoption it needs. |
The salvage is a reversal of order: program → evidence → facilitator system → stable materials spec → one packaged module → repeatable channel → additional modules → external marketplace. The original order maximized simultaneous unknowns.
Convert existing adult/continuing/workforce courses into governed AI-native learning systems with defensible oral and performance evidence — a course compiler plus governed evidence runtime, not "chat with your syllabus."
Offers: $15–30K Conversion Pilot (one bounded unit, 30–50 scored defenses) → $30–60K Production Course Conversion → $8–20K/course annual runtime.
Kinetic Data Lab: Circus Physics — eight 90-minute sessions, ages ~10–13, commodity materials. Synthesized from the kit murderboard: Stats Slam's measurement + Angle Arena minus lasers + FutureForge's capstone, run where the physical system is the curriculum.
Rules: EAC is a first-use lab, not portability proof; EEC written classification before any child cohort; a second site and second facilitator before provider-scale claims. $400–600/learner hypothesis.
Obligation-to-evidence reconciliation across special-education systems — not an IEP writer, and never a service-reduction engine. Red lines: no automated denial recommendations, no emotion recognition, no training on identifiable student data.
Staging: $15–30K offline pilot → $25–75K live implementation only when customer-funded and reusable → $25–50K/yr minimum.
Competency graph + evidence provenance + assistance-aware mastery and transfer checks with retention decay. Built only around workflows the paid wedges actually use — "infrastructure without a killer workflow becomes an elegant orphan."
| Gate | ProofWorks | Learning Lab | IEP Ledger |
|---|---|---|---|
| Day 30 | One paid Conversion Pilot closed or signed-in-procurement with a named owner and date | Lawful operating path confirmed, adult dry run done, 15 deposits (or documented smaller-cohort rationale) | Three design partners + one credible paid-scope candidate; synthetic/deidentified data only |
| Day 60 | Two paid pilots signed, one running; measure labor, reliability, corrections, expansion intent | Paid eight-week cohort underway; attendance, transfer, facilitator burden, safety | All tracks scored on the common card; provisional leader named |
| Day 90 | One unit completed and evaluated; seek $30–60K production conversion or second customer | All eight sessions complete; transfer + paid repeat demand; delayed-retention check scheduled | Build/park decision. Winner needs ≥60/100 + mandatory paid, legal, and economics gates → 70% of next quarter. Otherwise: no winner — park, restart, sell services, or stop |
Budget bands: ProofWorks-only validation ~$30–75K; adding a lawful Learning Lab cohort ~$75–175K — with EAC's facility, staff, and insurance counted economically even when they create no invoice. Explicitly prohibited in the first 90 days: custom tooling, production inventory, twelve concurrent modules, marketplaces, and outcome guarantees.
Learning is transfer, not task completion — measure after help is removed · a child using the product is not the buyer · a pilot is not evidence; a testimonial is not an outcome · district TAM is not obtainable revenue · hardware multiplies failure modes · no child data exhaust — never make surveillance the business model · human provenance can be a feature · every plan gets kill criteria; hope is not a milestone.