The market and evidence research behind Codex, built on your Ahura Core Technologies document. A team of local AI models running on our own machines read the papers and sources, and every quote and number was checked against its source before it went in.
Numbers show what each stretch added.
What a school sale is worth at real US district prices: 77.3 million US K-12, US college and England students × $2–15 per student a year (Kami, Pear Deck, Nearpod up to Khan Academy’s district price).
The ceiling if families paid Khanmigo’s $44 a year. Per teacher at SMART Lumio’s $59, it is ≈ $0.22bn. The pricing model moves the answer more than the headcount does.
Teachers already using an interactive whiteboard: England (government survey) and US union teachers (AFT survey). The display is rarely the constraint; who buys is.
2025 revenue change at the two listed display makers (Promethean’s parent, Boxlight), from their annual filings. A board maker is either a partner that carries Ahura’s sync or a competitor that builds it.
Market-research firms put the separate segments the MVP touches at $0.90bn to $46.2bn and disagree with each other 3–12×, so we show ranges, never one headline number.
Every MVP feature exists somewhere on its own; the loop does not. Cited answers (NotebookLM, Consensus, scite), annotation and peer comments (Hypothesis, Perusall, Glasp), text-to-audio (Speechify, NaturalReader), smartboard software (SMART Lumio, Promethean) and course analytics are all on sale. No product page we read shows a classroom board session saved into each student’s own study space, shared on to peers, with the teacher seeing where students struggle in the text.
Many states choose which reading screeners schools may use, and some pay for them. Taking Amira as the test case across 39 states and DC: the state pays in 4, it is on an approved list in 16, and it is left off in 7. Lists open on fixed windows and change (Alabama dropped Amira; several states switch lists from 2027–28). State price lists add screeners at $4–12.50 and reading software at $12–100 per student.
A review of 127 studies finds adaptive platforms generally improve outcomes; the value depends on transparency and the learner’s control.
A small gain in reading comprehension (d ≈ .24–.35); no reliable gain for fluent readers.
Captions, contiguity, signalling and segmenting are well evidenced; Ahura’s own video-to-text alignment is untested.
Backed by accessibility standards (WCAG 2.2, W3C COGA). Studies: no picture-or-text order always wins, cutting irrelevant text gives a small gain, line length showed no cost, dark mode no fatigue difference in one small study.
A 38-study review finds no evidence that dashboards alone raise achievement; one school study found gains when the dashboard led teachers to send feedback.
Social annotation is liked but studies are small; smartboard benefits depend on how teachers use them; three small studies of learning that carries on from class to home are positive, none randomized.
A chatbot grounded in a curated base was rated more accurate than ChatGPT 3.5 (4.78 vs 4.11 out of 5, 18 prompts). Course-built AI tutors beat no AI in two university trials. But no study compares grounded and general chatbots on learning, and a semester trial of a grounded chatbot found no gain.
No study tests annotation, peer sharing, smartboard sync or heatmaps together for learning. This is Codex’s claim to prove, with the proof of concept and a Clarke-center study.
| Decision | What it means for Codex |
|---|---|
| Your Core Technologies document is the basis | Market, competitors and the deck are built on the Codex MVP |
| Market shown both ways | Bottom-up and top-down side by side; the district price range leads |
| EU AI Act | Core Codex stays formative (no grading); grading later as a separate module; a legal read before any EU claim |
| Student data | Codex improves only on de-identified student data |
| Evidence | Grounding and node interoperability shown as a roadmap the proof of concept tests, not as proven |
| Order of proof | Clarke-center study in spring 2027, then district pilots |
Almost all of the work ran on local AI models on our own machines: they read the papers, searched and saved sources, drafted, and cross-checked each other. One Claude lead planned the questions, checked the results and wrote the memo. Every quote was matched to the saved source and every number recomputed by script before it went in; claims a model could not back up were dropped. Prices come from public district contracts and pricing pages, company figures from SEC annual filings, and market-report figures are labelled as such.