savvides · savvides · May 30, 2026 · May 30, 2026 · May 30, 2026
diff --git a/.github/ISSUE_TEMPLATE/case-study.md b/.github/ISSUE_TEMPLATE/case-study.md
@@ -0,0 +1,20 @@
+---
+name: Share a Case Study
+about: Share your experience applying the demand-validation framework at your institution or startup
+title: "[Case Study] "
+labels: case-study
+---
+
+**Institution or company type** (e.g., R1, community college, online program, K-12 district, corporate L&D, early-stage startup)
+
+
+**Journey phase(s) involved** (Pre-enroll, Apply, Onboard, Select & Enroll, Course Experience, Graduate & Beyond — see [data/higher-ed-jobs-atlas.md](https://github.com/savvides/edtechfounderstack/blob/main/data/higher-ed-jobs-atlas.md))
+
+
+**Which diagnostic questions did you apply?** (from [data/demand-validation.md](https://github.com/savvides/edtechfounderstack/blob/main/data/demand-validation.md))
+
+
+**What did you find?** (signal vs. noise — be specific)
+
+
+**Key takeaway**
diff --git a/.github/ISSUE_TEMPLATE/job-statement.md b/.github/ISSUE_TEMPLATE/job-statement.md
@@ -0,0 +1,23 @@
+---
+name: Suggest a Job Statement
+about: Propose a new validated job for the higher-ed jobs atlas
+title: "[Job Statement] "
+labels: job-statement
+---
+
+**Journey phase** (Pre-enroll, Apply, Onboard, Select & Enroll, Course Experience, Graduate & Beyond)
+
+
+**Person** (role, not title — e.g., "prospective student comparing programs," not "student")
+
+
+**Struggling moment** (what is the person trying to do, and what's blocking them?)
+
+
+**Current workarounds** (what do they do today, and why does it fail?)
+
+
+**Proposed job statement** (format: "When [situation], I want to [motivation], so I can [outcome]" — see [data/jtbd-interviews.md](https://github.com/savvides/edtechfounderstack/blob/main/data/jtbd-interviews.md))
+
+
+**Evidence source** (how did you validate this? interviews, data, institutional experience?)
@@ -15,8 +15,8 @@ All of it lives in `data/`, in three kinds of files.
 Structured domain knowledge, grounded in real sources rather than model training data:
 
 - **Regulatory** — FERPA, COPPA, and state privacy law (K-12); accreditation and accessibility (higher ed)
-- **Market** — competitive landscape by segment, buyer personas, funding landscape by stage, procurement, pilot benchmarks
-- **Frameworks** — ESSA evidence tiers, AI-native vs. bolted-on, the higher-ed jobs atlas, and founder traps
+- **Market** — competitive landscape by segment, buyer personas, buyer demand signals, funding landscape by stage, procurement, pilot benchmarks
+- **Frameworks** — ESSA evidence tiers, AI-native vs. bolted-on, defensibility moats, the higher-ed jobs atlas, founder traps, and the demand-validation toolkit (the 5-question diagnostic plus the JTBD Switch interview method)
 
 Each regulatory and market file carries a "last updated" date. Update cadence is roughly quarterly; regulatory data when laws change; the competitive landscape goes stale fastest.
 
@@ -26,7 +26,7 @@ Hundreds of peer-reviewed papers across the major learning-science topics, each
 
 ### Operator playbooks — `data/operator-lessons.md`
 
-Dozens of field lessons from operators and investors, distilled and attributed from the public archive of Lenny's Podcast and Lenny's Newsletter, then mapped to selling into schools, universities, and L&D. These are practitioner experience, not peer-reviewed evidence — the research corpus is the evidence layer, and the file says so.
+Dozens of field lessons from operators and investors, distilled and attributed from the public archive of Lenny's Podcast and Lenny's Newsletter, then mapped to selling into schools, universities, and L&D. These are practitioner experience, not peer-reviewed evidence — the research corpus is the evidence layer, and the file says so. The same practitioner-not-peer-reviewed labeling applies to the summit-sourced files (`data/buyer-demand-signals.md`, `data/ai-risk-and-trust.md`), which each carry their source and an evidence-tier note.
 
 ## How it's consumed
 

diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -1,5 +1,19 @@
 # Changelog
 
+## 2.1.0 (2026-05-30)
+
+### Demand-validation toolkit + knowledge from sibling repos
+
+Folds content from three sibling repos — Cracking Higher Ed (SXSW EDU 2026), the JTBD Switch toolkit, and ASU+GSV 2026 Summit Intelligence — into the knowledge base, centered on demand validation. Sources are CC BY 4.0 (Cracking Higher Ed, summit) and MIT (JTBD), attributed in-file; summit and JTBD material is labeled practitioner signal, not peer-reviewed.
+
+- **New `data/demand-validation.md`** — the 5-question demand diagnostic with scoring and the validation depth probes.
+- **New `data/jtbd-interviews.md`** — the JTBD "Switch" interview method (four forces, job stories, backward-timeline guide), reframed for edtech.
+- **New `data/defensibility-moats.md`** — the exposure spectrum, four moats, and the AI-substitution durability test.
+- **New `data/ai-risk-and-trust.md`** — AI's effect on learners and trust, with design responses founders can adopt.
+- **New `data/buyer-demand-signals.md`** — the durable jobs institutional buyers switch for, and how to read real demand.
+- **Fixed `data/higher-ed-jobs-atlas.md`** — completed from 11 to the full 15 jobs (the phase counts already implied 15).
+- Augmented `procurement-guide.md` and `pilot-benchmarks.md` with summit-sourced buyer and pilot realities; added case-study and job-statement issue templates; updated README, ARCHITECTURE.md, and CLAUDE.md.
+
 ## 2.0.0 (2026-05-29)
 
 ### Repositioned as a knowledge base

@@ -9,6 +9,10 @@ An open, AI-friendly knowledge base for edtech founders, built by ASU ScaleU. Th
 - `data/operator-lessons.md` — dozens of operator and investor lessons distilled from Lenny's Podcast and Lenny's Newsletter, mapped to edtech. Practitioner experience, not peer-reviewed; don't present it as research.
 - `data/ai-native-framework.md` — AI-native vs. bolted-on: criteria, the removal test, architecture patterns, pricing models, and the Karpathy hierarchy. Use it to classify a founder's AI posture.
 - `data/higher-ed-jobs-atlas.md` and `data/founder-traps.md` — ScaleU's SXSW EDU 2026 higher-ed framework: validated jobs across the student journey with saturation analysis, and the structural patterns founders miss.
+- `data/demand-validation.md` and `data/jtbd-interviews.md` — the demand-validation toolkit: the 5-question diagnostic with scoring and depth probes, and the JTBD Switch interview method for discovering and validating real demand.
+- `data/defensibility-moats.md` — how an edtech product stays defensible when LLMs can replicate features (exposure spectrum, four moats, the AI-substitution durability test).
+- `data/ai-risk-and-trust.md` — AI's effect on learners and trust, with design responses. Practitioner signals from the ASU+GSV 2026 summit, not peer-reviewed; don't present as research.
+- `data/buyer-demand-signals.md` — the durable jobs institutional buyers switch for. Practitioner signals, not peer-reviewed.
 - `ETHOS.md` — the seven principles, starting with "validate demand, not interest."
 
 Always cite the source: a named regulation, a paper's DOI, or the named operator.

@@ -30,12 +30,16 @@ Hundreds of peer-reviewed papers across the major learning-science topics: space
 - Higher-ed landscape, procurement, and accessibility
 - Funding landscape by stage — who funds edtech and what they require
 - Buyer personas, pilot benchmarks, and the ESSA evidence tiers (1–4)
+- **Buyer demand signals** — the durable jobs institutional buyers switch for ([`data/buyer-demand-signals.md`](data/buyer-demand-signals.md))
 
 ### ScaleU frameworks
 
-- **AI-native vs. bolted-on** — is your AI load-bearing or decorative ([`data/ai-native-framework.md`](data/ai-native-framework.md))
+- **Demand validation** — the 5-question diagnostic with scoring and depth probes ([`data/demand-validation.md`](data/demand-validation.md)), plus the JTBD Switch interview method for discovering real demand ([`data/jtbd-interviews.md`](data/jtbd-interviews.md))
 - **Higher-ed jobs atlas** — validated jobs across the student journey, with saturation analysis showing where everyone's already crowded ([`data/higher-ed-jobs-atlas.md`](data/higher-ed-jobs-atlas.md))
 - **Founder traps** — the structural patterns founders miss ([`data/founder-traps.md`](data/founder-traps.md))
+- **AI-native vs. bolted-on** — is your AI load-bearing or decorative ([`data/ai-native-framework.md`](data/ai-native-framework.md))
+- **Defensibility moats** — staying defensible when LLMs can copy your features ([`data/defensibility-moats.md`](data/defensibility-moats.md))
+- **AI risk & trust** — what AI does to learners before you ship a student-facing model ([`data/ai-risk-and-trust.md`](data/ai-risk-and-trust.md))
 - **The ethos** — seven principles, starting with "validate demand, not interest" ([`ETHOS.md`](ETHOS.md))
 
 ### Operator playbooks — `data/operator-lessons.md`

diff --git a/VERSION b/VERSION
@@ -1 +1 @@
-2.0.0
+2.1.0
@@ -0,0 +1,73 @@
+# AI risk and trust for edtech founders
+
+What AI does to learners, and to the trust between learners and the people around them, before you ship a student-facing model. For founders selling into K-12 districts, universities, and corporate L&D.
+
+*Note: these are practitioner signals from the ASU+GSV 2026 summit, not peer-reviewed evidence.* Where a claim is research-backed versus a panel assertion, this file says which. As of the ASU+GSV 2026 summit, the cognition question was the most contested theme of the week, and platform companies and practitioners were not in the same room on it.
+
+*Source: ASU+GSV 2026 Summit Intelligence — ScaleU. CC BY 4.0. Practitioner signals, not peer-reviewed research.*
+
+---
+
+## The honest caveat, up front
+
+The single most-cited number at the summit cuts against the hype. Ben Riley (Cognitive Resonance) pointed to Stanford SCALE's review of 800 studies of LLMs in education: only 20 showed causal impact, and virtually none of those were positive. So the research base is thin, and what causal evidence exists is not on the model's side yet. That is research-backed (it's a review of studies). Almost everything else below is panel assertion — operators reporting what they see, not controlled trials. Treat it as signal worth designing around, not proof.
+
+Riley's framing is the one to hold even if you reject his conclusion: a tool was deployed to roughly 500 million students before the longitudinal data exists. Wait for the RCTs to settle and you decide with a five-year lag. Deploy now without measurement and you become the data. Pick your error on purpose.
+
+---
+
+## The cognition and trust risks
+
+### Cognitive automation, not offloading (research-backed lean)
+Riley named the behavior "cognitive automation," not the gentler "cognitive offloading." He pointed to Carnegie Mellon and UCLA work showing "cognitive surrender" — students lose the ability to do the work once they try without the tool. The field datapoint: Sal Khan's own chief learning officer reported students typing "IDK" into Khanmigo rather than engaging with it. The risk isn't that students think less while using AI. It's that the capacity erodes. — Ben Riley (Cognitive Resonance)
+
+### Kids don't want the chatbot tutor (operator signal)
+Dan Meyer (Amplify) watches one benchmark: whether kids actually want to talk to a chatbot tutor. It has sat flat at roughly 5% for three years. His build rule follows from it — AI as an analytical layer for teachers, never direct-to-student. Joe Davis (KAIT Lab) goes the other direction with the same goal: AI-powered infrared pens that surface where students get stuck inside a problem set, so productive struggle stays the load-bearing part. — Dan Meyer (Amplify), Joe Davis (KAIT Lab)
+
+### Killing the butterfly (operator signal)
+Larry Berger (Amplify) gave the most evocative warning. The capabilities exist, he said, but every AI implementation he sees is "killing the butterfly" — the moment of collective wonder that pollinates the next thousand moments of learning. His board gave him six months to step back from running the company and figure out whether AI can keep the butterfly alive. He does not have an answer yet. That last part matters: a serious operator with every incentive to be bullish is openly unsure. — Larry Berger (Amplify)
+
+### Sycophancy and the praise problem (operator signal)
+Isabelle Hau (Stanford Accelerator for Learning) shared a stat from a visiting scholar: AI models praise children 13 times more often than humans do. Her read as a parent — if a model praises my child 13 times more than I do, kids start to expect it, and human-to-human relationships shift to match. The structural problem underneath: companies are incentivized to optimize for engagement, and sycophancy is a reliable way to get it. Your retention metric and the learner's development can point in opposite directions. — Isabelle Hau (Stanford Accelerator for Learning)
+
+### Anthropomorphism and developmental risk (operator signal)
+Scale makes this urgent. Prateek Maheshwari (Physics Wallah) runs mega-classrooms — 100,000 students in a single live AI session at $40 ARPU — and the student feedback keeps returning to one line: "AI is not judging us." That's the appeal and the hazard in one sentence. Matthew Biel (Georgetown pediatric psychiatry) frames these as "non-mutual transactional" relationships and warns that adolescent development requires rupture and repair, which a model that never judges and never pushes back cannot provide. Paul LeBlanc was blunter: "AI is going to make social media look like a day at the beach." — Prateek Maheshwari (Physics Wallah), Matthew Biel (Georgetown), Paul LeBlanc
+
+### Tools built for adults, handed to kids (operator signal)
+Anton Osika said Lovable hit $400M ARR in two years serving "the 99%" of non-developers, including nine-year-olds running real e-commerce sites. The underlying tools were not built for kids. Imagi exists as the safety wrapper for exactly that reason. If your product reaches minors — directly or because a teacher points it at a class — assume the base model was tuned for adults and the age-appropriateness layer is yours to build. — Anton Osika (Lovable)
+
+### Falling hope, rising anger (operator signal)
+Kevin Roose (NYT) and Casey Newton (Platformer) landed a different datapoint: a Gallup/Walton/GSV poll of 14-to-29-year-olds showed hope about AI down 9 points to 18% in one year, with a third of Gen Z AI users reporting anger. Garrett Lord coined the line that stuck — an "agency divide" between people who manage AI and people AI manages. The learners you serve are not uniformly excited. A meaningful share are anxious or angry, and they can tell which side of that divide a product puts them on. — Kevin Roose (NYT), Casey Newton (Platformer), Garrett Lord
+
+### The other side of the table (the optimist case, for honesty)
+The bull case came from James Donovan, head of learning and cognitive outcomes at OpenAI. His argument: the question isn't whether AI helps cognition but how the model is tuned. Model behavior elicits a human behavior, that behavior ladders up to cognitive outcomes over time, and tuned toward pedagogical alignment you get metacognitive gains. He pointed to a 20,000-student RCT in Estonia (University of Tartu and Stanford) as the model for generating real evidence. Note what this is and isn't: a stated thesis plus a study still running, not a result. Omar Abbosh (Pearson) gave the cleanest synthesis of both camps: "If you use it wrong, you will absolutely get dumber. If you use it right, you can get smarter." The institution's job — and your product's job — is enforcing the difference. — James Donovan (OpenAI), Omar Abbosh (Pearson)
+
+---
+
+## The design responses founders can adopt
+
+These came up at the summit as concrete moves, not theory. Each is a panel assertion about what's being built, not a proven outcome.
+
+### Refusal by design
+OMA Play's response to the developmental risk for the youngest learners is a screenless device for ages three to five with no face, that takes naps, shuts off at night, and refuses to engage 40% of the time on purpose. The design principle generalizes past toddlers: build friction in deliberately. The product question Hau's anthropomorphism work forces on every tutoring builder — what friction do you build in on purpose, and how do you measure when sycophancy is hurting the learner instead of just retaining them? — OMA Play
+
+### Age gates and an age-appropriate wrapper
+Imagi sits as the safety layer because the frontier tools underneath weren't built for kids. If your product touches minors, the age-appropriateness layer is a thing you build, not a thing you inherit. Expect district-grade equivalents for K-12 to be a category, not a feature.
+
+### Time caps and screen-time limits
+The OMA Play device caps engagement structurally — naps, nighttime shutoff, a designed-in refusal rate. Time limits are a design lever, not just a parental-controls afterthought. For a student-facing tutor, the cap is part of the pedagogy: it protects productive struggle and signals to a district buyer that you are not optimizing a child's screen time to the ceiling.
+
+### Keep the human as the user, when the evidence is thin
+Meyer's "analytical layer for teachers, never direct-to-student" is a posture, not a constraint you're stuck with. Until your own outcome data says otherwise, aiming the model at the teacher (where it surfaces who's stuck, drafts feedback, flags patterns) carries less developmental risk than aiming it at the child. It also clears district AI review faster.
+
+### The defensible institutional posture
+The summit's cleanest default for a buyer, which you can adopt as a product stance: assume model defaults push toward cognitive automation, not learning. Demand productive friction in the student-facing layer. Invest in teacher capability, not chatbot seats. Build the tool that makes the institution's enforcement job easier, not the one that quietly does the thinking for the student.
+
+---
+
+## Cross-links
+
+- The trust risk has a security twin. A student-facing model that touches FERPA records, reads untrusted content, and can send data out is the "lethal trifecta" — see the data-exfiltration point in [operator-lessons.md](operator-lessons.md). Prompt-injection filters top out around 97%, a failing grade, so architect the exfiltration path away rather than trusting a prompt to behave.
+- Whether AI is your product's load-bearing wall or a bolted-on chatbot changes how much of this risk you own. See [ai-native-framework.md](ai-native-framework.md). The deeper the AI sits in the product, the more the cognition and trust questions on this page are yours to answer, not the model lab's.
+
+*Last updated: 2026-05-30*