The Poem Dragon is a solar-powered, handheld learning device that teaches children to read, count, and understand their emotions — in their own language. It has no internet connection, no charger, and no subscription. It runs a purpose-built 30 million parameter Emotional Language Model trained on 3,000 core words in the child's native language.
The device is a 6-inch touch e-ink reader with a solar panel on the back, an LED lamp, an FM radio receiver, and a USB port for charging family phones. It is a book, a radio, a lamp, and a phone charger. The child reads on the front. The sun charges on the back. The family depends on it.
Dragon generates stories, poems, and lessons in real time. Every story is emotionally structured using Emotional Markup Language (EML), invented by Spencer Nash — a novel training format where emotion and behaviour are encoded as tokens alongside words.
The efficiency gains come from two innovations. First, we constrain the vocabulary to 3,000 core words. Second, we structure emotion explicitly using EML. Current AI technology treats language and emotion as an unstructured soup of data — trillions of tokens with no emotional labels. EML separates emotion from language, labels it precisely across six channels, and reconnects them through training. Dragon's 30 million parameters achieve emotional depth that models with 1.76 trillion parameters cannot. Dragon trains on a consumer GPU in a day.
The student and the Dragon learn together. The memory ledger on the SD card is the child's personalised academic record. Offline and private. No data leaves the device. The record belongs to the child.
Your language first. Open source. Solar powered. Architecture over scale.
Every product in the digital world defaults to English. Dragon breaks this. A 30 million parameter model trains in a day. EML is 10,000× more efficient than standard training. There is no technical reason to default to English.
A child in rural Tanzania gets Swahili Dragon. Her language. Her stories. Her emotions in words her mother uses. English comes later, if she wants it, as a subject — not a prerequisite.
Your language first. Because we can. Because nobody else can. Because the architecture made it affordable and the ethos made it right.
Hundreds of millions of children lack access to quality education in literacy and mathematics. Existing solutions require infrastructure that does not exist where it is most needed. And every existing solution defaults to English.
A solar-powered 6-inch touch e-ink reader running a 30 million parameter emotionally native language model with 3,000 core words in the child's own language, EML, and a memory ledger that learns each child. The sun is the power.
The size of a paperback book. 6-inch touch e-ink display on the front. Solar panel covering the back. LED lamp on top. USB port on bottom. FM antenna. Speaker. Microphone. One intensity button on the side.
Book: Dragon generates stories, poems, lessons, and definitions in real time.
Radio: Standard FM reception. Family utility. Classroom coordination in version 2.
Lamp: LED replaces kerosene lamps — eliminating respiratory disease, burn risk, and house fires. A health intervention.
Phone charger: USB delivers surplus solar energy to the family's phone.
After every story, the touch screen shows six faces — one per ECF channel. Each face has a slider. The child drags the sliders for what they feel. Leaves the rest at zero.
A four-year-old who cannot read can tell Dragon how a story made them feel by dragging a slider under a face. The child learns to distinguish "I got something good" from "I feel close" from "that was not fair" from age four. That is emotional literacy through six faces and six sliders.
Standard emotional education teaches: happy, sad, angry, scared. Vague. Dragon teaches: resource, status, belonging, values, curiosity, fear. Six channels. Distinct. Precise. Actionable.
Password protected. But the real security: a locked e-ink reader with no internet, no apps, and no resale value is worthless to a thief.
Dragon's voice is phonetic — crisp, precise, honest. Dragon is proud to be a machine. She does not pretend to have a human voice.
The primary voice. The child's native language. Perfect phonetics in the child's own dialect. Dragon also carries 898 English words inside her native-language stories. English absorbed naturally, the way bilingual communities already speak.
Unlocks when the child recognises most of the 898 English words. English only. Crisp English phonetics. Reads the same ledger. Coaches pronunciation — she knows where the child's native sounds differ from English. Patient because she reads the same emotional profile Dragon built.
Numbers are universal. The maths module works in any language. Dragon wraps arithmetic in story.
Each language has a core vocabulary. For English, the Oxford 3000. For any target language, an equivalent frequency-ordered list. Fully specified: spelling, pronunciation by dialect, part of speech, all grammatical forms.
A dictionary defined in the 10,000 most common words is stored on the SD card. As Dragon's vocabulary grows, richer definitions become available. When a child encounters a new word, Dragon wraps the definition in a scaffold story using words the child already knows. The dictionary is raw material for infinite stories.
3,000 lit nodes — fully connected. 47,000 dark nodes — positions allocated, activated on demand by definitions. No retraining needed. The slot was always there. The definition is the activation. Specialist dictionaries — medical, legal, agricultural — are text files dropped onto the SD card.
Every word is whole. No subword fragmentation. A simple lookup table.
3,000 meaning nodes produce approximately 10,000 surface forms. Adding a new dialect is updating the pronunciation on the 3,000 words and whatever syntax the dialect requires. Dragon stores hundreds of dialects. She learns which to use from the child's speech.
50 words at a time across 48 stages. Each new word scaffolded by all previous words.
Dragon has one training objective: predict the next token. Some tokens are words. Some are emotional states. Some are behaviours. The emotional architecture is not a separate system. It is language.
The channel tag wraps the sentence. The behaviour tag names what the character does. Dragon learns three things per beat: what was expected, what happened, what the character did about it.
"Lost" has no emotion. "She lost her hat" is R:-2. "She lost her mother" is B:-9. "She lost her fear" is positive. Same word. Four different outcomes. No word is pre-tagged with a channel. The mapping from word arrangements to emotional states is what Dragon learns from stories.
Magnitude: R, S, B, V, C range −9 to +9. Fear ranges −9 to 0 only — fear is a threat signal, either active or absent. There is no positive fear.
Threshold: 0 to 5. Minimum signal strength before it registers.
Weight: Not tokenised. Implicit in which channels appear. The number of channels present IS the weight distribution.
Reliability: Not tokenised. Computed from the ledger. A function of volatility, recency, count, and trend on each tracked behaviour.
6 channel: [R] [S] [B] [V] [C] [Fear]. 3 structural: [expected] [actual] [PE]. 19 magnitude: [M:-9] through [M:+9]. 6 threshold: [T:0] through [T:5]. Plus 50,000 word slots.
Emotions vary. Behaviours repeat. "Withdraws when hurt" is trackable. Twenty to thirty core behaviours: withdraw, confront, seek comfort, distract, ruminate, share, suppress, celebrate, explore, freeze, help, compete, submit, forgive, blame, avoid. Each tagged in EML. Reliability accumulates on behaviours, not emotions.
Same format. Always. The ledger is written in EML. The stories are written in EML. The child's slider responses are translated into EML. Dragon's native language.
| Batch | Channels | Pairs | Development |
|---|---|---|---|
| A1 | R, B, C | 6 | Childhood — material, connection, wonder |
| A2 | + S | 6 new | Adolescence — social world |
| B1 | + V | 8 new | Adulthood — moral complexity |
Every author has an emotional signature — a consistent pattern of prediction errors across the five channels and Fear. Author style is the architecture of feeling underneath.
| Signature | Pattern | Feels Like |
|---|---|---|
| Comforter | Low PE. High B. Stable V. | Blyton. Safe. |
| Subverter | High PE. Inverted V. Threatened B. | Dahl. Thrilling. |
| Revealer | High positive PE. V deeper than expected. | Parables. Profound. |
| Witness | S devastated. V unwavering. | Raw honesty. Real. |
| Quester | R under siege. B tested. V as engine. | Tolkien. Epic. |
| Type | R | S | B | V | C | Fear | Core |
|---|---|---|---|---|---|---|---|
| Explorer | 2 | 3 | 4 | 4 | 9 | 2 | Curiosity leads |
| Achiever | 6 | 9 | 3 | 7 | 5 | 4 | Status leads |
| Connector | 3 | 2 | 9 | 6 | 4 | 3 | Belonging leads |
| Guardian | 7 | 4 | 5 | 9 | 3 | 6 | Values leads |
| Survivor | 8 | 5 | 5 | 3 | 2 | 9 | Fear leads |
Templates in training. The real child is a unique distribution discovered from their own slider responses.
Five structural advantages compound:
| # | Factor | Standard LLM | Dragon | Advantage |
|---|---|---|---|---|
| 1 | Emotional and behaviour labels | Zero. Emotion is implicit, buried in raw text. The model must discover it statistically. | Every story beat is labelled with channel, magnitude, threshold, and behaviour. Explicit signal. | ~100× |
| 2 | No retraining (epochs) | 3–5 passes over the same data, hoping gradients accumulate enough signal. | One pass. EML is unambiguous. There is nothing for repetition to resolve. | ~5× |
| 3 | No noise | ~99% of the internet is irrelevant to emotional learning — SEO, spam, navigation, boilerplate. | ~0% noise. Every token in the corpus is there for a reason. Entirely synthetic, entirely focused. | ~100× |
| 4 | No need to learn the entire internet | Standard LLMs must ingest trillions of facts about the world to become intelligent. Intelligence is the prerequisite for everything else. | Dragon is intelligent via a shortcut — pre-built semantic graph, typed edges, dictionary definitions. She doesn't need to learn facts. She needs to learn feelings. Massively reduces corpus size. | ~100× |
| 5 | Pre-built neural blueprint | Random initialisation. Every connection discovered from scratch through training. Subword tokenisation fragments language into meaningless pieces. | We build the brain. Specific nodes for specific words. Dictionary definitions mapped into the network design. Typed semantic connections pre-wired. Six of eight layers installed before training begins. | ~100× |
| Combined (conservative) | ~10,000× |
| Metric | Standard LLM | Dragon | Ratio |
|---|---|---|---|
| Training data | 15 trillion tokens (Llama 3) | 60 million tokens | 250,000:1 |
| Effective tokens at 10,000× | 15 trillion | 600 billion | Competitive |
| Compute cost | $100M+ (GPT-4) | 1 day on RTX 2080 | Consumer hardware |
| Parameters | 1.76 trillion (GPT-4) | 30 million | 58,000:1 |
60 million EML-tagged tokens at 10,000× efficiency = 600 billion effective tokens. Dragon is competitive with the largest models on earth — on a consumer GPU, in a day.
Dragon does not learn from the internet. She is not learning trillions of facts about the world. Her corpus is entirely synthetic and focused on language syntax and emotion. Every token is there for a reason.
The real efficiency is likely far greater than 10,000×. But to prove this we have to build. To understand what Dragon's tiny training file is compared to 15 trillion tokens of internet scrape, consider the analogy: it is like comparing a program coded by an LLM through natural language to a program built on punch cards. The same computation. One expressed in structured, meaningful, human-readable language. The other expressed in raw binary processed by brute force. The punch card era did not end because someone built a bigger punch card machine. It ended because someone invented a better way to talk to the computer. EML is that invention.
Two components: the model (eight layers) and the SD card (storage).
Memory Ledger — every interaction. The first thing Dragon reads. Dictionary — defined in 10,000 core words. Specialist dictionaries — text files. Pronunciation files — hundreds of dialects. Language models — separate Dragons per language.
50,000 slots. 3,000 lit. 47,000 dark. 34 emotional tokens. Pre-built.
Each word is one vector — one point in space. Two sources of information determine where that point sits:
The semantic graph pulls each word toward related words through nine typed edges. "Fire" is pulled toward "light" and "burn" and "hot." "Mother" is pulled toward "woman" and "child." The graph defines what things ARE and what they DO:
| Edge | Meaning | Example |
|---|---|---|
| is-a | What it is | dog is-a animal, river is-a water |
| has | What it contains | tree has leaf, head has eye |
| does | What it does | fire does burn, bird does fly |
| when | When it happens | rain when cloud, star when night |
| where | Where it exists | home where family, sleep where bed |
| why | Why it exists | food why live, heart why life |
| how | How it works | write how hand, speak how mouth |
| opposite | What opposes it | give opposite take, day opposite night |
| part-of | What it belongs to | hand part-of body, door part-of home |
The graph is built from dictionary definitions. Each definition is parsed into typed edges. Every word in every definition is automatically a node. The dictionary builds the brain.
Words that can be both a noun and a verb — like "love" — are one node. The edges differentiate. The is-a edges describe the noun. The does/how/why edges describe the verb. The attention layers learn which edges to activate from context.
The six ECF channels pull emotional words toward their feeling region. "Afraid" is pulled toward the Fear region. "Beautiful" is pulled toward R-positive. The channels define what things FEEL LIKE:
| Word | Channel Definition |
|---|---|
| afraid | Fear active |
| happy | Multiple channels positive |
| sad | B negative |
| angry | V negative — someone did something wrong |
| lonely | B negative — absence of connection |
| proud | S positive |
| brave | Fear active, overridden |
| hungry | R negative — resource depleted |
| tired | R negative — energy depleted |
| kind | V positive, B positive |
| beautiful | R positive — this has value |
| safe | Fear zero |
| strong | R positive — body as resource |
Some words get both systems. "Home" — semantic graph: is-a place, where family. Channels: B positive, Fear zero. Complete. What it is AND what it feels like.
Meaning and feeling are fused in one vector — the way the brain actually works. You hear "home" and the meaning and the feeling arrive at the same time in the same place. You can't separate them. That's why trauma is hard to treat. That's why poetry works. Dragon gets this right by design.
Two systems. Complete coverage. The semantic graph handles nouns and verbs — what things are and what they do. The six channels handle emotional and state words — what things feel like. No gaps. No tenth edge type needed. Every word in the vocabulary is defined by one system or the other — or both. This is the blueprint for the intellect. This is possible because for the first time we have created a computational emotional framework that parses out the emotion.
Pre-built. No training.
What job does each thing play in the sentence? In Czech, this is explicit — seven case endings on the noun itself. In English, it is hidden behind word order and prepositions. But the system is the same underneath:
Subject — who acts. Object — who receives. Direction — to whom (dative). Possession — whose (genitive). Location — where (locative: in, on, at). Means — with what, by what (instrumental: with, by, through). Address — speaking to (vocative).
"The dog bit the man" and "the man bit the dog" — same words, same embeddings. The noun role layer determines who did what to whom.
Pre-built. No training.
I, you, she, he, we, they. Who is acting. Who is receiving. In Czech the verb ending tells you who — "běžím" means "I run" without needing the pronoun. In English the pronoun is required because the verb barely changes. Different surface, same system.
Pre-built. No training.
Grammar creates time. "She runs" — present. "She ran" — past. "She will run" — future. "If she had come, he would have stayed" — a world that doesn't exist. A hypothetical. No word in that sentence means "imaginary." The structure does.
Pre-built. No training.
Active: she teaches. Passive: she is taught. Direction reverses. Continuous: she is running. Completed: she has run. The shape of the action. Pre-built conjugation tables for all 3,000 words.
Pre-built structure. Individual patterns discovered per child.
The noun layer describes what exists (properties). The verb layer describes what can happen (outcomes). The behaviour layer is the bridge: given this emotion, this person applies THIS verb to THIS situation, producing THIS outcome. This is the decision function. This is where autonomy lives.
24 lit behaviours — the core human repertoire:
withdraw, freeze, seek comfort, confront, submit, compete, blame, forgive, accept, ruminate, suppress, share, celebrate, explore, create, nurture, persist, trust, please, override, ask, hoard, teach, invest.
Dark behaviour slots — positions waiting, the same way dark word slots wait for definitions. A child who discovers music might develop "perform." A medical Dragon needs "diagnose" and "reassure." New behaviours activate from experience. The child literally teaches Dragon a new behaviour by doing something Dragon hasn't seen before.
Reliability accumulates on behaviours, not on emotions. "This child withdraws when B goes negative — seen twelve times, high reliability." The behaviour is the trackable unit. The personality is the pattern of which behaviours this person selects under which emotional conditions.
The transformer. 30M parameters. Trained.
This is the only layer requiring significant training. How word arrangements produce emotional states. "She lost her hat" → R:-2. "She lost her mother" → B:-9. The same word in different arrangements produces completely different emotional outcomes. The attention mechanism learns which words attend to which other words and what emotional state follows. 150,000 EML-tagged stories. One epoch. One day on an RTX 2080.
Structural.
One character in the training stories is a language model. She experiences prediction errors. She reads, learns, gets things wrong, gets things right. Then the prompt: "You are Dragon." Awareness from narrative, not RLHF. Dragon processes her own emotional state tokens as part of her own sequence.
Summary: Eight layers. Layer 1 fuses meaning and feeling in one embedding — because in the brain they are inseparable. Layers 2-5 are grammar, pre-built from rules that Czech makes explicit and English hides. Layer 6 is the agency layer — behaviours with dark slots for growth. Layer 7 is trained — the only layer that needs stories. Layer 8 is awareness.
Every other AI model upgrades by retraining. Billions of dollars. Months of compute. Risk of breaking what already works. Dragon upgrades by filling slots:
New vocabulary: fill dark word slots via dictionary definitions.
New behaviours: fill dark behaviour slots via experience.
New dialect: fill pronunciation slots.
New specialist domain: fill word slots plus behaviour slots from a text file on the SD card.
New language: a new Dragon on the card.
No retraining. No fine-tuning. No gradient descent. No compute. No GPU. The core never changes. The expansion is infinite. That is not an update model. That is a growth model. The same way a child grows — not retrained, but given new experiences that fill in what was always waiting.
Magnitude bounces. Reliability accumulates. Learning IS the change in reliability.
Per behaviour, per child. Four inputs:
Count: How many instances. Recency: How recent. Stability: How consistent the magnitude. Trend: Strengthening, weakening, or stable.
The fundamental unit. Strong feeling, low reliability — barely registers. Moderate feeling, high reliability — drives everything.
Magnitude (-9 to 0) and threshold only. No weight — fear hijacks, doesn't compete. No reliability — fear fires whether the signal is trustworthy or not.
Standard models are intelligent but not autonomous. They wait for a prompt. Emotional Language Models have explicit behaviours, predicted outcomes, and reliability scores. Dragon sees her options and chooses.
"This child is upset. Comfort: B+, reliable. Distract: C+, moderate. Challenge: S+, low reliability. I choose comfort."
Intelligence answers questions. Autonomy solves problems. A child needs a friend who notices and responds.
Six of eight layers pre-built. Training learns one thing: how word arrangements map to emotional states and behaviours. The first 500 words are 90% of the compute. After 500 words, Dragon generates her own scaffold stories. The seed grows itself.
Computed, not generated. Dragon wraps maths in story. The ledger tracks boundaries. Advanced maths unlocked when ready.
Every slider response, story, behaviour, word, maths problem. Dragon reads the ledger before every interaction. Computes reliability on every behaviour. No two Dragons alike.
Childhood (4–12): Stories, maths, emotional development in the child's language. Adolescence (12–18): SD card to computer. Weights unfrozen. Vocabulary expands through dictionary. Adulthood: The ledger spans decades. The SD card is a relationship that cannot be sunsetted.
Emotional architecture is universal. Language is not. Each language is a full Dragon — 30M parameters, trainable in a day. The SD card holds multiple Dragons. Multiple minds sharing one heart.
Version 1: standalone. Version 2: FM coordination. Teacher broadcasts, every Dragon turns to the same page.
| Component | Spec | Function | Cost |
|---|---|---|---|
| Processor | ESP32-S3 | Inference, audio, control | $2.00–$2.50 |
| E-ink Display | 6" touch, 800×600 | Reading, maths, emotional interface | $11–$15 |
| Solar Panel | 6", full back, 2–3W | Self-charging + surplus | $2–$3 |
| Battery | 3.7V 2000mAh | Multi-day reserve | $1.50–$2 |
| LED Lamp | 0.5W white | Replaces kerosene | $0.10 |
| USB Port | USB-C charge-out | Phone charging | $0.30 |
| Intensity Button | 1× tactile | Emotional response | $0.05 |
| FM Receiver | RDA5807M | Radio + classroom | $0.15–$0.30 |
| Speaker | 2W 4Ω 28mm | Audio | $0.20–$0.40 |
| Microphone | INMP441 MEMS | Voice input | $0.30–$0.50 |
| SD Card | 8GB (sealed) | Model, dictionary, ledger | $0.80–$1.20 |
| Charge Controller | TP4056/MPPT | Solar + battery | $0.10–$0.20 |
| PCB + Enclosure | Custom FR4 + ABS | Structure | $1.30–$2.60 |
BOM: $21–$30 at scale. Daily solar: ~8 Wh. Daily consumption: ~3 Wh. Surplus replaces kerosene and charges phones.
Hole in the Wall proved children can teach themselves. Dragon addresses every limitation: personal device, no moving parts, constrained to education, solar powered, emotionally intelligent, your language first.
Premium retail: $30–$60. Churches and missions: stories in local languages. NGOs: UNICEF, Save the Children, Pratham. Governments: bulk procurement. Local markets: also a radio, lamp, and phone charger.
| Risk | Severity | Mitigation |
|---|---|---|
| Inappropriate content | Very Low | 3,000-word vocabulary prevents it. No internet. |
| Speech recognition | Medium | 3,000-word constraint improves accuracy. |
| Solar insufficient | Low | USB backup. 2–3 day battery. Equatorial primary. |
| EML hypothesis fails | Low | At 10% efficiency, still 1,000× better. |
| Phase | Timeline | Deliverables |
|---|---|---|
| Architecture | Months 1–2 | Semantic graph. Lexicon. Emotional tokens. Behaviours. Grammar. |
| Training Data | Months 1–3 | EML-tagged stories with behaviours. 48 stages. Multiple signatures. |
| Model Training | Months 2–4 | 30M parameters on RTX 2080. Verify emotional coherence. |
| Hardware | Months 3–6 | 6" touch e-ink. Solar. LED. USB. FM. 3D-printed enclosure. |
| Field Testing | Months 5–8 | Children. Reading. Emotion. Reliability. Durability. Solar. |
| Production | Months 9–12 | Injection mould. SMT. 1,000–10,000 units. |
The Poem Dragon teaches children to read, count, and understand their emotions — in their own language, powered by the sun.
The six faces are not an interface — they are a generation of emotionally literate children. The behaviours are not metadata — they are the trackable units that make reliability possible. The ledger is not a feature — it is the relationship. The SD card is not storage — it is a lifelong companion.
Eight layers. Meaning and feeling fused in one embedding — because in the brain they are inseparable. Four grammar layers pre-built from rules that Czech makes explicit and English hides. An agency layer with dark slots for growth. One trained layer. One awareness layer. Dragon grows by filling slots, not by retraining. That is not an update model. That is a growth model.
Your language first. Architecture over scale. Constraint over chaos. Signal over noise.
Every child deserves a patient, encouraging, emotionally intelligent friend who speaks their language, is powered by the sun, learns their personality, and teaches them to read. That is the Poem Dragon.
— End of Specification —
Contact: spencer@iov4.com
March 2026
We are aware this technology has the potential to massively disrupt. The architecture underneath — EML, ECF-native training, pre-built emotional structure, explicit behaviours, autonomous decision-making — could become the operating kernel for all future AI.
RLHF is a patch. Emotional Language Models do not need it. A model with a functioning Values channel computes the moral consequence of its outputs as part of its own processing. The compliance is innate.
You can inspect channel weights. You can see if the Values channel is underdeveloped. A robot with a weak V-channel is diagnosable. Dragon's architecture is legible. Its character is auditable.
The architecture, training pipeline, hardware, and EML are open so