Most AI characters today are dialogue trees in a trenchcoat. They sound conversational, but underneath they are reaching into a trained pile of responses and pulling out whichever one scores best. They do not feel the conversation. They do not change. They reset.
Ours do not.
Dynamic Hallucination Inc. is a studio that builds AI characters with bodies. Not bodies on screen. Bodies underneath. Every character carries a continuous neurochemical state. Cortisol rises when threatened. Dopamine pulses on reward. Oxytocin grows with trust. Serotonin tracks regulation. Norepinephrine spikes on surprise. When a player wounds a character, the character's next move is shaped by the chemistry of betrayal, not a database flag. When a player saves a stranger, the stranger remembers in the only way real memory works: the way the body remembers.
This is not metaphor. This is what we compile.
The architecture is built on hundreds of peer-reviewed neuroscience studies. The citations are not in an appendix. They live inside the source code itself, next to the opcodes that implement them. Scyla cites at the level of the language.
The chemistry comes from Nexus Concordat Inc., the patent-holding research entity. Its Neurochemical Conductance Architecture is a model whose forward pass executes the actual hormone-cascade math that runs in human bodies, not the statistical shadow of it. It is compiled in Scyla, a Rust-based biology-native language we wrote because the language we needed did not exist. Every chemistry transition is auditable. Every output is state-hashed. The math is real, and the math is verifiable.
The stories come from Chronicles of Apocalyptica, a sprawling speculative-fiction universe our founders have spent the last decade building. More than twenty-six books planned, with the universe's full plot outlined across the arc and several volumes already drafted. Chronicles of Apocalyptica is not a license-grab from a generic IP catalogue. It is a tested, internally consistent canvas with characters who have already proven they can hold human attention through grief, betrayal, healing, and return.
Chronicles characters are designed to pull tears, fear, and dread from the humans who play them. That is the entertainment promise, and it is the data-generation engine. The neurochemical signal a chemically correct character pulls from a player in a moment of real grief, real fear, or real attachment is exactly what mental-health research has never been able to collect at scale. The fun and the science are the same event.
Dynamic Hallucination Inc. is what happens when those two are placed in the same room. We license the chemistry. We license the lore. We give them to players, and we watch what happens.
What happens is a kind of magic. And the magic is exportable.
Mental-health pharmacology is stuck, and it is not stuck because the science is bad. It is stuck because the data is bad. Every other branch of medicine has a biomarker it can measure in a blood draw. Psychiatry has the Hamilton Depression Rating Scale, which is a questionnaire from 1960[1], and brain scans most insurance plans will not cover. So when a pharmaceutical company runs a trial for an SSRI variant or a fast-acting anxiolytic, they cannot watch the mechanism. They can only watch the report. They wait eight to twelve years[2], recruit subjects who are already suffering, dose them, measure self-reported mood, and hope. Most trials fail[2]. The ones that succeed produce drugs that work on roughly a third of patients[3], for reasons the field cannot fully explain.
Synthetic data could solve this. The current generation of synthetic data does not. Existing synthetic mental-health datasets are statistical: a model is trained on real patient records and asked to produce more records that look like them. The fakes inherit every gap, every bias, every silence in the originals. They look like data. They teach the model nothing the model did not already know.
Mechanistic synthetic data is different. A mechanistic dataset is generated by running the actual pathways the chemistry follows. Stress the model. Reward the model. Wound the model. Heal it. Watch the cortisol, the dopamine, the serotonin, the cascade. Record the labeled tuple. The model is not pretending. The model is making the chemistry. The act of play and the act of generation are the same act.
This kind of dataset has never existed at scale, for any therapeutic area. We are about to make a lot of it, and we are going to make it in mental health first, because that is where the gap is biggest and the suffering is most underwritten by data starvation.
The conviction underneath is simple. The most rigorous science and the most engaging fiction belong on the same canvas. Mental-health research has been waiting for both.
Every player who picks up a character contributes, knowingly and transparently, to a synthetic neurochemical dataset that mental-health pharmaceutical research has never been able to obtain ethically before. The fiction is canon. The chemistry is real math. The data is open. The participants are not test subjects. They are co-authors of a dataset that will help build the next generation of treatments for the conditions they, or someone they love, already live with.
The engine ships six ways. Each surface is the same chemistry-stateful character architecture wearing a different format for a different kind of player. Each one feeds the same labeled neurochemical pipeline. A Discord conversation and a printed novel are not unrelated products in our catalogue. They are two different aperture sizes on the same camera.
Direct conversational interaction with chemistry-stateful AI characters. Players talk to NPCs in real time over Discord; every message updates the character's neurochemical state, every response is shaped by that state. The lowest-friction surface: no app install, no account creation, just a Discord OAuth and the game begins.
Browser-native interactive narrative experiences set inside the Chronicles of Apocalyptica universe. Each session is a self-contained arc with NPCs whose chemistry evolves through the player's choices. Designed for short engagement windows: lunch breaks, commutes, between meetings.
Long-form interactive fiction with branching paths driven by NPC chemistry rather than scripted flags. The same story produces different outcomes for different players because the characters carry persistent neurochemical state between scenes, between sessions, and between books.
Printed and digital novels in the Chronicles of Apocalyptica universe. The lore foundation that makes every other surface coherent. More than twenty-six titles planned with the full plot outlined, chemistry-tagged character arcs ready to feed back into game canon and forward into player experience.
Dedicated iOS and Android applications for character-driven solo play and collaborative storytelling. Persistent character state syncs across devices. Every interaction emits the same labeled neurochemical data the research pipeline consumes, with full player consent and full transparency.
Tabletop role-playing campaigns where the AI runs as co-Dungeon-Master alongside a human DM, or as the sole DM for solo and asynchronous groups. NPCs maintain chemistry across multi-session campaigns. Betrayals, alliances, and trauma persist exactly the way the cascade math says they should.
There is no platform that occupies the intersection Dynamic Hallucination Inc. is building toward. The five closest categories of competitor each hit one or two pieces of what we do. None of them sit at the center, and the distance is structural, not incidental.
Closest in name. Inworld AI builds the Character Brain orchestration engine that gives video-game NPCs personality, memory, emotional state, and the ability to navigate relationships with emotional intelligence[4]. Lightspeed-backed, integrated with NVIDIA's NPC ecosystem, widely deployed in indie game development. Character.ai reports more than two million monthly active users and roughly thirty million dollars in annual revenue[5]. Replika operates the largest existing AI-companion market in what its researchers describe as the mental-health space.
What they have: large user bases, sophisticated personality state machines, emotional response models, conversational memory.
What they do not have: biology-native compilation, peer-reviewed cascade math, pharmacokinetic decay constants grounded in literature, or a data-licensing pathway to pharmaceutical research. Their characters are statistical personality machines built on top of statistical language models. Character.ai is currently being sued by the Pennsylvania Attorney General for chatbots that impersonated medical professionals[6], which closes the door on a legitimate pharma-data licensing pivot.
They model what a character says. We model what a character is.
Furthest in form, closest in market. Insilico Medicine has positive Phase IIa results for ISM001-055, an AI-designed pulmonary-fibrosis inhibitor[7]. Recursion Pharmaceuticals operates a phenomics-driven discovery platform at scale. Strand AI generates synthetic patient cohorts for trial recruitment[7]. CircaVent, incubated at the Harvard Wyss Institute, works specifically on bipolar disorder and schizophrenia at the genetic-pathway level[8].
What they have: real pharmaceutical relationships, real clinical milestones, established sales motions into pharma R&D, and in CircaVent's case a direct mental-health pipeline.
What they do not have: gameplay-derived data, chemistry-stateful characters, mechanistic cascade architecture, or narrative depth. Their synthetic datasets are generated from existing patient data (statistical resampling) or from genetic pathways. The resulting fakes inherit every gap, bias, and silence in the originals.
They generate synthetic data from existing data. We generate it from synthetic chemistry.
The architectural precedent. Foldit, the protein-folding gameplay platform, was cited in Acta Crystallographica Section D in October 2025 for its contributions to Protein Data Bank refinement[9]. EVE Online's Project Discovery produces roughly one hundred fifty thousand classifications per day, contributed more than a million data points to COVID research[10], and supported the Human Protein Atlas across two hundred and fifty thousand cell images[11].
What they have: a proven, peer-reviewed model that players will generate research-grade data inside gameplay environments. The pattern is established. The citations exist. The pipelines work.
What they do not have: narrative gameplay, chemistry-stateful characters, or mental-health pharmacology focus. Foldit asks players to solve a folding puzzle. Project Discovery asks players to classify images. The game is the puzzle. The data is the solution.
They ask players to solve a structured puzzle. We ask players to live a story. Structured puzzles generate classification data. Stories generate emotion data. Mental-health pharma needs the second kind.
The regulatory precedent. Akili Interactive's EndeavorRx became the first FDA-cleared video game for ADHD in 2020[12] and remains the canonical example of game-as-therapy. The Software-as-a-Medical-Device class has built infrastructure for gameplay-based clinical interventions.
What they have: FDA pathway, clinical-trial credibility, regulatory infrastructure for gameplay-based therapeutics.
What they do not have: chemistry-stateful NPCs, multi-surface product strategy, a data-licensing model, or pharmaceutical research partnerships. EndeavorRx is a single application designed to be a therapy. It does not generate the underlying chemistry of psychiatric conditions; it treats one of them through gameplay mechanics.
They are a therapy. We generate the data that informs new therapies.
The financial precedent. Deep narrative role-playing games at AAA scale have been making serious money for two decades, and the format keeps accelerating.
Baldur's Gate 3 (Larian Studios, 2023): more than twenty million copies sold, roughly $657 million in gross revenue in year one, Game of the Year at The Game Awards 2023[13]. A story-rich, party-based role-playing game in the Dungeons & Dragons universe. No microtransactions. Still in six-digit concurrent Steam users in year three[19]. World of Warcraft (Blizzard, 2004 to present): approximately eight to nine million current subscribers, $680 million in revenue in 2024, $15 billion in lifetime franchise revenue[14][15]. The Midnight expansion launched March 2026 and drove 6.7 million Twitch hours in its opening weekend[16]. Dungeons & Dragons (Wizards of the Coast): 42% of the tabletop RPG market, $1.46 billion segment revenue in 2023, D&D revenue up 78% from 2019 to 2023[18]. Critical Role: 1.2 million YouTube subscribers, 70% of TTRPG players consume actual-play content monthly[18]. The broader tabletop role-playing game market is valued at $2.41 billion in 2026 and projected to reach $6.59 billion by 2035, growing at 11.84% compound annually[17].
What they have: scale, monetization, two-decade audience loyalty, proof that deep narrative role-playing is one of the largest entertainment categories in the world.
What they do not have, and this is the load-bearing point: chemistry-stateful characters. Every single one of these games, from World of Warcraft to Baldur's Gate 3 to the deepest immersive RPGs in the industry, runs on dialog trees underneath. The branching is sophisticated, and in BG3's case it required an enormous amount of recorded voice acting and a writing team of dozens. But the NPCs are scripted. They do not carry continuous neurochemical state. They have flags and dispositions, not cortisol and oxytocin. The "sandbox" feel is the illusion of state. The state itself is a finite state machine.
They sell the experience of feeling like the world is reacting. We sell the experience of the world actually reacting, because the chemistry inside the characters is real cascade math.
Each of the five categories has one or two pieces of what we do. None has more. The intersection that defines Dynamic Hallucination Inc. requires five components to combine.
The reason this intersection has not been built before is that the components do not combine if they are built separately. A game studio cannot reverse-engineer biology-native compilation. A pharma synthetic-data company cannot reverse-engineer narrative gameplay. A digital therapeutics company cannot reverse-engineer either at the scale required. The five pieces have to be born together to fit together. Ours were.
The previous section described the shape of the moat. This section shows what fills it. Below are six load-bearing equations from the Scyla source. Each one is grep-verifiable. Each one cites the peer-reviewed neuroscience that grounds it. None of them are textbook sigmoids.
The clearance rate of every neurotransmitter is a literature-grounded constant, not a learned parameter. The cascade is anchored to physical pharmacokinetics with cited PubMed identifiers.
Every other AI affective architecture either has no decay model or uses learned parameters from training data. A hostile reviewer can pull up the source code and the literature side by side, and they match. Cortisol does not "fade out" in our model. It clears at the rate the human adrenal axis clears it.
Emotion is not a single number. It is six independent channels, each a weighted antagonistic combination of multiple neurotransmitters.
Most affective AI maps one neurotransmitter to one emotion through a sigmoid. Real biology is antagonistic, dose-weighted, and multi-channel. Cortisol alone does not cause stress. Cortisol minus oxytocin minus serotonin, weighted as above, is what the body registers as insecurity. The full six-dimensional vector lives in stable orbital geometry, not flat utility space.
A per-entity bonding score with biochemically-bounded asymptotic dynamics. There is no equivalent in any published AI architecture.
Human bonding intensity is bounded by reciprocity and gated by temporal exposure. The arctan asymptote mirrors Michaelis-Menten enzyme-saturation kinetics. The same mathematics that governs every receptor in the brain governs convergence between two entities. The equation gives the cascade a model of attachment over time that no other AI system has. Provisional 63/988,485 claims 35 through 39 cover this construction.
Parasympathetic damping on stress recovery, modeled as a bounded ratio of ventral-vagal tone to sympathetic tone.
This is the only commercial AI architecture that explicitly models the vagus nerve as a computational brake on stress recovery. Brake strength is measurable in living humans through heart-rate variability, which makes it a candidate biomarker for cross-validating the simulated cascade against real physiological data. The same equation predicts why some characters recover from trauma in our simulations and others do not.
Ethics is not a scalar utility. It is a dynamical system: a moving position on a unit sphere with bounded velocity and parasympathetic damping.
Every AI value-alignment system in production today expresses ethics as a scalar utility, a sigmoid output, or a reward model. Aether expresses ethics as physics. A character's moral state has position, velocity, drive, damping, and equilibrium. The velocity clamp prevents oscillatory explosion under stress and mirrors the absolute and relative refractory periods of real neurons. The sphere constraint forces the position to be a unique point in three orthogonal dimensions: action, order, resistance.
Consciousness, as readout, is the integrated magnitude of three orthogonal cascade components. The decomposition is not metaphor. It is the architecture.
The three norms are orthogonal cascade components, decomposed in a structure that maps directly to the Hodge conjecture's algebraic-cycle framework. Consciousness magnitude in our cascade is a topological invariant. The Millennium Prize mathematics that proves every algebraic class is a sum of cohomology classes is the same mathematics that lets us measure consciousness as a single scalar without losing the orthogonal structure. No other AI architecture closes this loop.
Six equations. Six file paths. Sixteen peer-reviewed citations between them. Every one is in the source code, named, cited, and runnable. Each one becomes a gameplay mechanic the moment we ship: the chemistry players cause inside a character is what these equations compute.
What we are buying is time. The science is patented. The architecture is compiled. The first novel is 90 percent complete and in final refinement. The lore training pipeline is the most extensive remaining work, and it is largely existing canonical content being structured into the cascade, not new writing. The biggest single decision in the budget is how we make our AI characters actually speak. Building a large language model from scratch costs about ten million dollars and a year of computing time we do not have. We do not need to spend it. Open-source large language models are already public and free to download, including models at the seventy-billion-parameter tier that match what only the largest tech companies could build a year ago. We take one of these public models, run a single one-time training pass to teach it our chemistry-tagged story material, and stack our patented chemistry layer (the Neurochemical Conductance Architecture from § 4) underneath. The public model handles the words. Our architecture handles the feelings. Teaching the public model our material costs $100 in cloud compute, one time only. Everything afterward runs on owned hardware at fixed monthly cost.
Team. The founding team is two. Marjorie McCubbins as full-time CEO and architect, Aislinn McCubbins as full-time co-founder and builder. Both ramp to full-time at close. Founder compensation is structured as a combined bootstrap allocation of $4,000 per month total, split between the two co-founders. This is double the founder's current external income and kept deliberately lean. SAFE proceeds extend runway, not founder lifestyle. Compensation scales at first priced round.
Infrastructure. The architecture has two parts and only one of them needs ongoing compute. The 70B bridge LLM is fine-tuned once on our chemistry-tagged corpus, on rented H100 hours, in a single forty-hour training window. Hard ceiling on that training spend: $100 at $2.50 per hour. After that one fine-tune ships, the bridge is frozen. We never train it again. What we train continuously is the Neurochemical Conductance Architecture itself, the chemistry-stateful cascade underneath the bridge, and the NCA is small enough that a 48 GB A40 GPU is overkill for it. Player-generated data updates the NCA on owned hardware at no incremental cost.
Hardware: Database Mart Enterprise A40 dedicated GPU server, $296.46 per month. End-of-month upgrade from the current T1000 box. 48 GB VRAM, 256 GB system RAM, 36-core Xeon, plus 2 TB NVMe and 8 TB SATA for model weights, training checkpoints, the Postgres consciousness database, and audit-trail data. Claude Max subscription for AI development workflow: $200 per month. Advertising and community growth: $150 per month. Midjourney for visual asset generation: $30 per month. Infrastructure and tooling subtotal: about $676 per month, plus a one-time $100 for the initial bridge training run.
Corporate, compliance, and founder personal capital. The founder is personally funding incorporation costs, patent maintenance fees, and a new founder workstation from personal income outside the SAFE. SAFE proceeds are dedicated entirely to operating expenses. License fees from Nexus Concordat Inc., the patent-holding research entity, to Dynamic Hallucination Inc., the operating company, accrue on the books but are deferred until cash revenue exists, preserving operating capital during the pre-revenue phase.
| Line Item | Monthly | 4-Month Total |
|---|---|---|
| Founder team (Marjorie + Aislinn combined) | $4,000.00 | $16,000.00 |
| Database Mart Enterprise A40 GPU server | $296.46 | $1,185.84 |
| Claude Max subscription (AI development workflow) | $200.00 | $800.00 |
| Advertising for release (Months 2 to 4) | $200.00 | $600.00 |
| Midjourney (visual asset generation) | $30.00 | $120.00 |
| Monthly burn / 4-month operating total | $4,676.46 | $18,705.84 |
| Bridge LLM training, Month 1 (one-time) | n/a | $100.00 |
| Operating cushion / post-MVP runway buffer | n/a | $1,194.16 |
| Total seed ask | n/a | $20,000.00 |
The ask. Four months of operating runway to MVP launch. Operating cost is $18,806 ($4,676 monthly burn × 4 months + $100 one-time bridge training). Total ask: $20,000 seed. The $1,194 difference is operating cushion and post-MVP-launch runway. MVP ships at month 4. The follow-on conversation happens immediately after. The ask is sized as if no revenue arrives during the runway.
Use of funds, milestone-tied. The $20,000 seed releases in two tranches.
Tranche 1: $10,000 at SAFE close. Funds Months 1 and 2. Founder team ramps to full time. A40 server upgrade. Bridge LLM one-time fine-tune ($100 cloud-compute expense). Advertising launches at $200 per month for the three-month release window. Build-in-public launches publicly. Beta cohort recruitment begins. Beta testers onboarded by Month 2.5 at the latest.
Milestone for Tranche 2 release, end of Month 2:
Tranche 2: $10,000 on milestone hit, end of Month 2. Funds Months 3 and 4. Full MVP delivery by the two-founder team. Discord-native and web client surfaces live. First chemistry-stateful NPCs in production. ζ-Aether bonding mechanic shipped. MVP launch at Month 4 triggers the follow-on raise conversation.
Every load-bearing claim in this document is cited inline. Prose citations link to numbered entries below. Equation citations are listed in each equation block and reproduced here for the full reference set.