Status: research briefing that grounds V9_PRODUCT_ANALYSIS.md,
V9_features.md, and V9_ARCHITECTURE.md. Date compiled: 2026-06-13. Audience:
product, learning-science, generative-AI, science, and research leads scoping
V9.
This document captures the external state of the art for AI-powered educational
content generation and "learn anything in a fun way" products as of mid-2026.
Every architecture decision in V9_ARCHITECTURE.md cites a row here. The repo's
"wire real data, no stubs" rule applies: these techniques are to be
implemented and composed, not name-dropped.
Cutoff note. The authoring model's knowledge cutoff is January 2026; this briefing was researched in June 2026. Items dated after January 2026 are flagged [POST-CUTOFF]. Claims that could not be verified against a primary source are flagged [UNVERIFIED] and must not be cited as fact. Several developments materially change V9's design relative to a Jan-2026 view — most importantly the shutdown of OpenAI's Sora app (do not build on it) and the consumer launch of DeepMind's Project Genie.
1. Source-grounded study assistants & multi-format generation#
The decisive 2026 lesson: Socratic chat tutoring is commoditized; the differentiation has moved to grounded, multi-format, interactive generation.
| Technique / product | Source | What V9 takes from it |
|---|---|---|
| NotebookLM — source-grounded generation into Audio Overviews (interruptible), Video Overviews, Mind Maps, flashcards + quizzes, infographics, slide decks; outputs restricted to your sources, which is the antidote to hallucination. [POST-CUTOFF] 2026 adds: Cinematic Video Overviews (Mar 2026), 10 infographic styles (Mar 20 2026), grounded flashcards/quizzes with cross-session progress (Mar 2026), default Gemini 3.5 + secure code execution (Jun 2026). | blog.google / workspaceupdates.googleblog.com (2026) | The grounding-first principle (Aletheia G1) and the multi-format output model (text, audio, film, explorable). V9's wedge beyond NotebookLM: embodiment, computed simulation, and a mastery loop — NotebookLM summarizes your docs; V9 guides you through reality. |
| Gemini Guided Learning (LearnLM lineage, folded into Gemini 2.5; launched Aug 6 2025) — Socratic step-by-step tutoring with multimodal output (images, diagrams, videos, interactive quizzes). LearnLM is no longer a standalone model. | ai.google.dev/gemini-api/docs/learnlm ; blog.google (Aug 2025) | Confirms Socratic + multimodal is table-stakes; pedagogically-tuned models exist. V9 differentiates on grounding + embodiment + interactivity, not on "we have a tutor mode." |
| ChatGPT Study Mode (Jul 29 2025) — Socratic scaffolding across Free/Plus/Pro/Team. Key weakness: runs on general ChatGPT, not vetted sources, so it can present fabricated material. | technologyreview.com (Jul 2025); openai.com/index/chatgpt-study-mode | The cautionary tale that makes the Aletheia gate non-negotiable: an ungrounded tutor will fabricate, and that is fatal for a learning product. Grounding is the moat. |
| Claude Learning Mode / Claude for Education (Apr 2 2025; opened to all users Aug 2025) — Socratic "guide discovery, don't answer"; Claude Artifacts produce runnable interactive widgets in-chat. | anthropic.com/news/introducing-claude-for-education | The generative-explorable primitive (Artifacts) as the cheap path to on-demand interactivity (Hephaestus, §4), plus the integrity-mode discipline (Metis already has teach/hint/practice/do-not-complete). |
| Khanmigo / Khan Academy — real mastery curriculum + GPT-4 coach. [POST-CUTOFF] Khan publicly admitted only ~15% of students with access use Khanmigo, ran Oct 2025–Apr 2026 tests, and is shipping a 2026 redesign reoriented around motivation, not just mastery. | edtechinnovationhub.com (2026); blog.khanacademy.org (Jan 7 2026) | The engagement constraint is the binding one, not model quality. V9's game/awe/story craft is the structural answer to exactly the failure Khan named. |
Net rule for V9: ground everything (NotebookLM's trust), tutor Socratically (commodity, but required), and win on the layer no incumbent unifies — embodied + computed-interactive + mastery-tracked + awe-driven.
2. Embodied teachers & accurate explainer video#
The critical constraint: cinematic generative video models cannot reliably render accurate on-screen text, equations, or diagrams, and produce only short clips. For real teaching, the viable paths are (a) script-driven avatars and (b) programmatic precise animation.
| Technique / product | Source | What V9 takes from it |
|---|---|---|
| Synthesia 3.0 (Oct 1 2025) — Express-2 full-body avatars; Video Agents (real-time two-way interactive avatars). [POST-CUTOFF] $200M Series E at $4B (Jan 26 2026) funding the agentic rollout. Avatar reads your script ⇒ no fabricated facts. | synthesia.io/post/synthesia-3-0 ; cnbc.com (Jan 26 2026) | Validates the embodied-presenter path for Chiron. V9's advantage: the script is grounded+computed (Aletheia), and the avatar is Oshun's own Psyche stack (cross-session memory, emotion-aware), not a SaaS talking head. |
| HeyGen — LiveAvatar (real-time WebRTC avatar, BYO-LLM); Avatar V (studio-quality from ~15s of footage, performance/appearance separated). [POST-CUTOFF] Avatar V API ~$0.05/s, Custom Motion, HyperFrames (Apr–May 2026); legacy Interactive Avatar sunset Mar 31 2026. | heygen.com/blog (2026) | The real-time conversational-tutor pattern (Chiron live), and the proof that low-latency embodied dialogue is production-ready in 2026. |
| D-ID V4 Expressive Visual Agents [POST-CUTOFF] (Mar 16 2026) — up to 4K, sub-0.5s latency, LLM-driven expressions, inline interactive UI (charts, forms, quizzes) overlaid on the avatar. | d-id.com/news (Mar 16 2026) | The interactive-overlay pattern: an embodied teacher with live charts/quizzes embedded — directly informs Chiron + Hephaestus integration. |
| Cinematic models — illustrative, not explanatory. Veo 3.1 (Oct 15 2025; [POST-CUTOFF] 4K + free 720p tier ~Apr 2026) is best-in-class for narrated B-roll with native audio. Runway Gen-4.5 (Dec 2025) for consistency. Sora 2 (Sept 30 2025) — [POST-CUTOFF] app shut down Apr 26 2026, API winding down Sept 24 2026; Disney deal collapsed. DO NOT build on Sora. Pika = social, not teaching. | developers.googleblog.com ; variety.com / en.wikipedia.org (Sora shutdown, 2026) | Cinematic models are a component for illustrative scenes only (provenance-stamped), never the teacher or the source of facts/equations. The Sora shutdown is a direct architectural warning: do not anchor a teaching product on a single cinematic video vendor. |
| Programmatic precise animation — Manim + LLM-to-Manim (the only path to accurate on-screen math/physics). TheoremExplainAgent (arXiv 2502.19400, ACL 2025 oral) plans+codes 5-min Manim explainers (o3-mini 93.8% on TheoremExplainBench/240 theorems). [POST-CUTOFF] Math-To-Manim agentic pipeline on "Claude Mythos"; ManiBench (arXiv 2603.13251, ~Feb 2026) measures visual-logic drift / syntactic hallucination; a 2026 training paper reports Qwen3-Coder-30B at 94% render-success [UNVERIFIED figures]. | arxiv.org/abs/2502.19400 ; manim.community ; github.com/HarleyCoops/Math-To-Manim | The computed-correct explainer-film path (V9 §9): for math/physics, generate Manim from the verified Kalika derivation, not from free text. Reliability (render-success, visual-logic drift) is the known open problem ⇒ a validation/self-repair loop is mandatory (the Aletheia + G5 discipline). |
Net rule for V9: Chiron is a script-driven, grounded, embodied teacher (Psyche/Synthesia-class), not a cinematic-video hallucination; STEM visuals are computed (Kalika→Manim, Nyx render), and cinematic clips are illustrative B-roll only, vendor-diversified.
3. Adaptive learning, ITS, and knowledge tracing#
Production SOTA is the decades-old, deployed-at-scale mastery paradigms (now with an LLM tutor bolted on), plus an active attention/Transformer + LLM knowledge- tracing research frontier.
| Technique / product | Source | What V9 takes from it |
|---|---|---|
| Mastery engines in production — ALEKS (Knowledge Space Theory; ALEKS-for-Calculus Sep 15 2025), Carnegie Learning MATHia (ACT-R cognitive tutor; CODiE 2025 winner Oct 22 2025), DreamBox, Duolingo Birdbrain (IRT/half-life regression + GPT-4 features). | mheducation.com ; carnegielearning.com ; blog.duolingo.com/birdbrain | The mastery-engine-does-sequencing, LLM-does-conversation split. V9's Mnemosyne plays the mastery/sequencing role; Chiron plays the conversational role. Duolingo is the engagement bar to clear. |
| Knowledge tracing lineage — BKT (Corbett & Anderson 1994) → DKT (Piech 2015, LSTM) → attention/Transformer KT: AKT (KDD 2020, arXiv 2007.12324), SAINT/SAINT+ (Riiid, EdNet, arXiv 2010.12042). Critical caveat: pyKT (NeurIPS 2022, arXiv 2206.11460) showed many deep-KT gains shrink under fair evaluation — AKT/DKT remain strong baselines; cross-paper AUCs are not comparable. | arxiv.org/abs/2007.12324 ; arxiv.org/abs/2206.11460 | V9's mastery model uses a proven KT family (AKT-class) as the backbone, evaluated honestly (pyKT discipline), not a fragile bleeding-edge claim. Mnemosyne already implements BKT/DKT-style tracing. |
| 2025–26 KT frontier — LLM-representation KT: NTKT (next-token KT, arXiv 2511.02599, v2 Jan 26 2026) for cold-start generalization; HISE-KT (arXiv 2511.15191) for explainable KT; FlatFormer (arXiv 2512.06629, Dec 2025) efficient flat Transformer with forgetting-curve bias [UNVERIFIED self-reported SOTA]. | arxiv.org/abs/2511.02599 ; arxiv.org/abs/2511.15191 | The cold-start path (a new learner's first wonders) benefits from LLM-representation KT and from content-semantic embeddings (Sophia/Nous) — adopt cautiously, behind the AKT baseline, with honest evaluation. |
| Strongest causal efficacy evidence — human-AI hybrid. Stanford Tutor CoPilot RCT (2024–25): +4 p.p. mastery (+9 p.p. for students of lower-rated tutors) at ~$20/tutor/yr. Fully-autonomous LLM tutors lack comparable causal evidence. | nssa.stanford.edu/studies/tutor-copilot | V9's "escalate to a human" path (Metis already supports human/hybrid tutor types) is evidence-based, not a fallback afterthought; the strongest results are hybrid. |
Net rule for V9: a proven mastery/KT backbone (Mnemosyne, AKT-class, honestly evaluated) drives sequencing and the flow channel; the LLM (Chiron) does conversation; the human is an evidence-backed escalation, not a stub.
4. Generative interactive simulations, explorables & world models#
The "fun, interactive learning" frontier bifurcated by mid-2026 into (a) cheap, on-demand LLM-generated explorables and (b) heavyweight world models — with strong evidence that interactivity beats passivity for learning.
| Technique / product | Source | What V9 takes from it |
|---|---|---|
| Explorable explanations lineage — Bret Victor "Explorable Explanations" (2011) + "Ladder of Abstraction"; Nicky Case (explorabl.es, Parable of the Polygons); Distill.pub (dormant since 2021). The modern shift: LLMs make explorables generatable on demand (Claude Artifacts, ChatGPT Canvas → p5.js/three.js/React widgets). | worrydream.com/ExplorableExplanations ; explorabl.es ; distill.pub | The explorable-on-demand primitive is Hephaestus's cheap path: for any idea without a pre-built kernel, generate a runnable, verified interactive widget. This is "the most under-exploited fun-learning primitive." |
| Peer-reviewed prompt-to-sim efficacy — a 2026 physics-education study (Phys. Rev. PER; arXiv 2509.22929) found students who generated their own simulations via AI prompting reached conceptual understanding equal to using pre-built sims, and both digital methods beat physical labs (on electric potentials). | arxiv.org/pdf/2509.22929 | The strongest academic validation of V9's interactive pillar: generated, manipulable sims genuinely teach. Justifies investing in Hephaestus's generative-explorable path, not just hand-built kernels. |
| PhET Interactive Simulations (CU Boulder, ~159 sims, billion+ uses) — the correctness gold standard, but a finite hand-built library with no generative layer. | phet.colorado.edu | The correctness bar for explorables. V9's edge: Nyx/Kalika make our explorables computed and correct (PhET-quality accuracy) and generative (beyond a finite library). |
| World models — DeepMind Genie 3 (Aug 5 2025, 720p/24fps, minutes-consistent navigable worlds); [POST-CUTOFF] Project Genie consumer launch Jan 29 2026 (AI Ultra, 60-sec sessions). Decart Oasis 3 (Jun 10 2026, hours of photorealistic driving). World Labs Marble (persistent downloadable 3D worlds; $1B raise Feb 2026). | deepmind.google/blog/genie-3 ; the-decoder.com (Jan 29 2026); worldlabs.ai | World models are a P2/P3 horizon for immersive explorables, not a P1 dependency: they are not accurate enough for STEM facts/equations and sessions are short. V9's P1 explorables are computed (Nyx/Kalika) + generative-widget, which are correct today. |
Net rule for V9: P1 interactivity = computed kernels (Nyx/Kalika) + verified generative widgets (correct and on-demand); world models are a later immersive layer, never the source of factual accuracy.
5. Learning-science backing (canonical, citable)#
V9's pedagogy implements established research, most of which Mnemosyne already
encodes (libs/mnemosyne/). All citations independently verified against
primary/DOI records; caveats are stated, not buried.
| Principle | Canonical source | What V9 does with it | Honest caveat |
|---|---|---|---|
| Bloom's 2-sigma | Bloom (1984), Educational Researcher 13(6), DOI:10.3102/0013189X013006004 | Aspiration for one-to-one mastery tutoring at near-zero marginal cost (Chiron + Mnemosyne). | Never replicated at full 2σ; real tutoring meta-effects ≈ 0.3–0.4σ (Nickow/Oreopoulos/Quan 2020). Cite as aspiration, not promise. |
| Mastery learning | Bloom (1968), "Learning for Mastery"; Kulik (1990) meta-analysis | Mnemosyne mastery gates: advance on demonstrated mastery, time variable. | ~+0.5σ in meta-analysis; ~1σ in Bloom's framing. |
| Spaced repetition | Ebbinghaus (1885); Cepeda et al. (2006), Psych. Bulletin 132(3) | FSRS/SM-2 scheduling in libs/mnemosyne/core — review just before forgetting. |
Robust across 1,300+ effect sizes. |
| Retrieval practice / testing effect | Roediger & Karpicke (2006), Psych. Science 17(3), DOI:10.1111/j.1467-9280.2006.01693.x | Every lesson ends in active recall, not re-reading (G3, the Mnemosyne checkpoint). | 1-week recall ~61% (tested) vs ~40% (restudy). |
| Cognitive load theory | Sweller (1988), Cognitive Science 12(2), DOI:10.1207/s15516709cog1202_4 | Prometheus scopes each lesson to the prerequisite frontier (Atlas) to minimize extraneous load. | Intrinsic/extraneous/germane triad is from the 1990s (Sweller/van Merriënboer/Paas 1998), not the 1988 paper. |
| Desirable difficulties | Bjork (1994); Bjork & Bjork (2011) | Interactives withhold answers until the learner predicts; spacing/interleaving by design. | "Desirable" only if the learner has the background — hence Atlas scoping. |
| Flow | Csikszentmihalyi (1990), Flow | Mnemosyne knowledge-tracing tunes difficulty into the challenge-skill flow channel. | Flow needs high challenge matched by high skill (between anxiety and boredom). |
| Awe & learning | Keltner & Haidt (2003), Cognition & Emotion 17(2), DOI:10.1080/02699930302297; Keltner (2023), Awe | The Theia/beauty pillars as a deliberate awe→curiosity engine. | Awe → curiosity/openness is supported; awe → measured learning gains is suggestive, not RCT-proven. Design for it; don't oversell. |
6. What the landscape tells us about V9's scope#
- No incumbent unifies the five pillars. NotebookLM grounds but doesn't embody or simulate; Synthesia/HeyGen embody but don't ground or simulate; PhET simulates but doesn't generate; Khan tutors but can't engage; world models generate worlds but aren't accurate. V9's novelty is composition, exactly as V8's novelty was composition.
- Engagement and trust are the binding constraints, not model quality. Khan's ~15% admission (trust-of-attention) and Study Mode's fabrication risk (trust-of-truth) are the two failures V9's two structural bets (delight + the Aletheia gate) directly target.
- Grounding + computation is the moat. The only durable defense against the fabrication that plagues open tutors is binding claims to sources (Sophia) and computing values with real kernels (Nyx/Kalika) — capabilities the monorepo already has and a wrapper-of-an-LLM competitor does not.
- Build P1 on what is correct today. Computed kernels + verified generative widgets + script-driven embodied avatars are production-ready and accurate in 2026; world models and cinematic video are illustrative/immersive layers for later, never the factual core. The Sora shutdown is the cautionary tale for over-anchoring on a single generative-video vendor.
Bottom line: every component V9 needs has a published, working 2026 precedent, and most exist in the Oshun monorepo already. The work is integration, grounding, and the consumer experience — not invention.
Sources#
AI study assistants & grounded generation
- NotebookLM 2026 features — https://workspaceupdates.googleblog.com/2026/03/new-ways-to-customize-and-interact-with-your-content-in-NotebookLM.html ; https://techcrunch.com/2026/06/08/notebooklms-new-update-will-help-you-build-source-repository-from-chat/
- Gemini Guided Learning / LearnLM — https://blog.google/outreach-initiatives/education/guided-learning/ ; https://ai.google.dev/gemini-api/docs/learnlm
- ChatGPT Study Mode — https://www.technologyreview.com/2025/07/29/1120801/openai-is-launching-a-version-of-chatgpt-for-college-students/ ; https://openai.com/index/chatgpt-study-mode/
- Claude for Education — https://www.anthropic.com/news/introducing-claude-for-education
- Khanmigo engagement / 2026 redesign — https://www.edtechinnovationhub.com/news/only-15-percent-of-students-with-access-to-khanmigo-actually-use-it-khan-academy-admits ; https://blog.khanacademy.org/khan-academy-reimagined-for-districts-2026/
Embodied teachers & explainer video
- Synthesia 3.0 / Series E — https://www.synthesia.io/post/synthesia-3-0 ; https://www.cnbc.com/2026/01/26/nvidia-alphabet-vc-arms-back-synthesia.html
- HeyGen 2026 — https://www.heygen.com/blog/heygen-may-2026-release ; https://www.heygen.com/blog/announcing-avatar-v
- D-ID V4 — https://www.d-id.com/news/v4-expressive-visual-agents-real-time-llm-connected-interaction/
- Veo 3.1 — https://developers.googleblog.com/introducing-veo-3-1-and-new-creative-capabilities-in-the-gemini-api/
- Sora 2 shutdown — https://en.wikipedia.org/wiki/Sora_(text-to-video_model) ; https://variety.com/2026/digital/news/openai-shutting-down-sora-video-disney-1236698277/
- TheoremExplainAgent — https://arxiv.org/abs/2502.19400 ; Manim — https://www.manim.community/ ; Math-To-Manim — https://github.com/HarleyCoops/Math-To-Manim ; ManiBench — https://arxiv.org/html/2603.13251 ; Manimator — https://arxiv.org/html/2507.14306v1
Adaptive learning, ITS, knowledge tracing
- ALEKS — https://www.mheducation.com/about-us/news-insights/press-releases/mcgraw-hill-releases-ai-powered-aleks-for-calculus.html ; https://www.aleks.com/about_aleks/knowledge_space_theory
- Carnegie Learning — https://www.businesswire.com/news/home/20251022158855/en/Carnegie-Learning-Wins-CODiE-Awards-for-MATHia-and-Patterns
- Duolingo Birdbrain — https://blog.duolingo.com/birdbrain/
- AKT — https://arxiv.org/abs/2007.12324 ; SAINT+ — https://arxiv.org/abs/2010.12042 ; pyKT — https://arxiv.org/abs/2206.11460
- NTKT — https://arxiv.org/abs/2511.02599 ; HISE-KT — https://arxiv.org/abs/2511.15191 ; FlatFormer — https://arxiv.org/abs/2512.06629
- Tutor CoPilot — https://nssa.stanford.edu/studies/tutor-copilot-human-ai-approach-scaling-real-time-expertise
Interactive sims, explorables, world models
- Explorable Explanations — https://worrydream.com/ExplorableExplanations/ ; https://explorabl.es/ ; https://distill.pub/2020/communicating-with-interactive-articles/
- Prompt-to-sim physics efficacy — https://arxiv.org/pdf/2509.22929
- PhET — https://phet.colorado.edu/
- Genie 3 / Project Genie — https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/ ; https://the-decoder.com/google-deepmind-opens-project-genie-to-us-subscribers-for-real-time-ai-world-generation/
- Decart Oasis 3 — https://techcrunch.com/2026/06/10/decarts-new-world-model-can-simulate-hours-of-photorealistic-driving-with-some-caveats/
- World Labs Marble — https://www.worldlabs.ai/blog
Learning science
- Bloom 2σ — https://doi.org/10.3102/0013189X013006004 ; 2-sigma critique — https://www.educationnext.org/two-sigma-tutoring-separating-science-fiction-from-science-fact/
- Mastery learning — https://eric.ed.gov/?id=ED053419
- Retrieval practice — https://doi.org/10.1111/j.1467-9280.2006.01693.x
- Cognitive load — https://doi.org/10.1207/s15516709cog1202_4
- Flow — Csikszentmihalyi (1990), Flow, ISBN 9780060162535
- Awe — https://doi.org/10.1080/02699930302297 ; Keltner (2023), Awe, ISBN 9781984879684
Verification caveats#
- Independently verified (primary/official sources): Sora 2 shutdown (Apr 26 2026 app / Sept 24 2026 API); Synthesia 3.0 + Series E; D-ID V4; HeyGen 2026 releases; NotebookLM 2026 features; ChatGPT Study Mode; Claude for Education; Gemini Guided Learning; Genie 3 + Project Genie; Decart Oasis 3; TheoremExplainAgent; and all pedagogy + KT baseline citations.
- Flagged [UNVERIFIED] — do not present as fact: FlatFormer/DenoiseKT self-reported SOTA; the 2026 Manim training-paper render-success figures; Squirrel AI 2025–26 corporate/efficacy specifics; any "Genie 4"; exact Veo 3.1 4K/length specs; HeyGen "Avatar 5"/15-second-training claim. Several 2026 arXiv IDs (26xx prefixes) are recent and were not all fetched in full.
- This briefing executed no pipeline; product-side capability judgments are
cross-referenced to
V9_GAP_ANALYSIS.md(source reads), not to checkbox ledgers.