The training plan is the TTRPG plan.

What it is

A tabletop roleplaying game where mental health is a mechanic, not a topic. Players, or psychiatry residents in training, or therapists running structured sessions, create characters in the Chronicles of Apocalyptica universe. Each character has a full D&D-style sheet AND a current neurochemical state panel: a 10-dimensional vector across cortisol, dopamine, serotonin, norepinephrine, oxytocin, GABA, acetylcholine, endorphins, glutamate, melatonin. The state is set by rolling, by clinical profile, or by GM design.

The character then plays in that state. Aether is the GM. Encounters are designed for psychiatric exploration: triggers, social pressure, relational tension, threat, loss, reward. The character responds in their chemistry, not as a generic NPC. Sessions are persistent. The character who was in a PTSD state last Tuesday is in the same PTSD state next Tuesday unless something in the world or the session changes it. State is canonical, not transient.

Chemistry biases behavior. It does not determine it. A character in PTSD state is still themselves, just colored by hypervigilance and irritability. A character in mania state is still themselves, just amplified and impulsive. Identity, memory, relationships, current chemistry, and current situation are the five inputs to the behavior generator. The character emerges from the integration. The chemistry colors. It never erases.

The proof, in one image

This is what the architecture does, today, with the cardiac substrate already running:

The chemistry colored the response without erasing the character. That is the architectural claim made operational.

Cardiac is the existence proof that the training pipeline works. A specialist cell trained end-to-end, signed, verifiable against a public registry, serving requests in production. Apply the same pipeline to a different corpus, get a different specialist. psych_gen1, training today on a public psychiatric corpus, is the second proof point.

The three products hiding in one architecture

One invention, three distinct things it enables. Naming them separately so they don't collapse into each other.

Product 1 · The Engine

A persistent character-state engine. Five inputs (identity, memory, relationships, current chemistry, current situation) feed a deterministic substrate that produces a character's behavior. Chemistry biases. It does not determine. Same five inputs, same response, every time. This is the invention. Everything else depends on it.

Product 2 · The Game

Chronicles of Apocalyptica. A TTRPG where Aether is the GM, characters live in canon lore, and the new mechanic is that chemistry is a stat on the sheet. Already running. Years of world-building. This is the first commercial application. The game funds the research. The game is also the safe simulation environment, because it is a game.

Product 3 · The Test Bed

A future vertical, not the headline. Because the engine is deterministic and replayable, it can serve as a substrate for testing things the field has no other way to test: drug-effect direction predictions, scale-validation experiments, training scenarios with objective ground truth. Downstream of the engine. Downstream of the game. A door the architecture opens, not the product we are selling today.

What is operational today · receipts

Each item below is independently verifiable. URLs serve live content; SHA-256 verification recipes are published on the cell-demo page.

• cardiac_gen3 specialist cell. ~32 million parameters, chemistry-conditioned through the 34-head cascade, trained end-to-end on a 20-million-token curated cardiology corpus. Public registry entry at /registry/cardiac-registry.json. Live inference at /cell-demo.html.
• Signed inference manifest per response. Each emitted token is bound to a manifest containing cell_id, opcode-tree hash, sample parameters, and SHA-256 over the manifest body. External parties can recompute the hash against the public registry and confirm byte-exact provenance without access to model weights.
• Vanilla transformer baseline. An apples-to-apples standard PyTorch transformer trained on the identical corpus, serving alongside the gen3 cell so evaluators can compare emission patterns under matched seed, temperature, and top-k. Same data, no cascade, no manifest path.
• Cardiac → Qwen 2.5 14B language bridge. The cardiac cell emits substrate tokens; a Qwen 2.5 14B language model translates them into readable English under a constrained system prompt that prohibits invented biology. Operational at POST /cardiac-14b/sample. End-to-end latency 6 to 8 seconds.
• Per-character clinical pipeline. A synthetic patient profile (selected from seven canonical diagnostic templates) is bound to a character record, mapped through documented marker-to-hormone reference equations into a 10-dimensional neurochemical state vector, then sampled through the cardiac → 14B bridge to produce state-grounded output. Persisted to chronicles.character_cardiac_log with full audit trail.
• Public corpus pull infrastructure. Currently active across PubMed E-utilities (targeted MeSH queries), OpenFDA (drug labels + FAERS adverse events), ChEMBL (molecule-target-mechanism), ClinicalTrials.gov (mental-health arms), NIMH RDoC. Each pull writes a manifest with source URL, query string verbatim, retrieval timestamp (UTC), license attribution, byte count, and SHA-256 per output file. Forty-plus manifests already written and on disk.
• Reproducibility discipline. Raw data and cleaned data are stored in separate directories. The corpus-purpose document, CORPUS_PURPOSE.md, declares the training I/O contract and source weighting in writing, before training runs.

What has shipped since the last update

The training plan, in one piece

Three cells deep is the trajectory. Each cell is a step on the same pipeline.

• cardiac_gen3 is live. The existence proof. Pipeline produces a working chemistry-conditioned cell end-to-end. The 34-head cascade, the signed manifest, the Merkle DAG lineage, the cell-as-opcode tree, all operational. Queryable at /cell-demo.html right now.
• psych_gen1 is training right now. The generalization proof. Same architecture, same pipeline, same hyperparameters, different corpus (psychiatric vocabulary, drug pharmacokinetics, adverse-event signal). Vanilla baseline running on the local A5000; Scyla 4-brain cascade staged for the next pass. First checkpoint expected within hours.
• A third cell is planned. Domain selected based on what makes the generalization story unarguable. Endocrine, neurology, GI, whichever closes the cleanest argument. Two cells is a coincidence. Three cells is a method.

The patent fortress was designed for exactly this shape. Cell-as-opcode-tree covers any specialist composition. The signed-manifest claim covers any cell's inference verifiability. The substrate-independent memory weighting claim covers any cell's pharmacokinetic decay handling. Adding specialists does not widen the patent surface; it deepens the existing claims with operational evidence. Every new cell strengthens the same four-claim portfolio.

Where the trained cells land · inside the TTRPG

The cells are the substrate Chronicles of Apocalyptica runs on. That's the connection.

In the game, every character has a neurochemical state panel as part of their sheet. When Aether narrates that character's behavior, the response is generated through the trained specialist cell that matches the character's chemistry context, then translated through the language layer (Qwen 2.5 14B, currently) into readable narrative. A character with elevated cortisol and low serotonin gets a response colored by what the psychiatric cell has actually learned about cortisol-and-low-serotonin states, drawn from manifest-documented public corpus.

The training plan is the game's spine. Without trained cells the TTRPG runs on whatever default the LLM produces. Fine for play, not grounded in biology. With trained cells the same game becomes a substrate-grounded simulation: the same character with the same starting state responds consistently across sessions, and a perturbation (missed dose, sleep loss, stressor) drifts behavior in directions the cell has actually learned. That consistency is what makes the game a research artifact in addition to a product. The Xena demo above is what this looks like in practice.

cardiac_gen3 is in production as the proof that this loop runs end-to-end. psych_gen1 is what makes the loop work for the psychiatric scenes that are the heart of the game.

What this means for Mayo

The Mayo conversation is the same TTRPG deployed to a different audience.

Mayo's psychiatry residency program needs simulated patients. The current state of practice is standardized-patient actors: hired humans trained to perform clinical presentations. Expensive, inconsistent across sessions, cannot be reset to a specific neurochemical baseline on demand. What the TTRPG already does, persistent character state, chemistry-grounded responses, replayable across sessions, audit-trail-clean, is exactly what a residency training program needs and cannot get from human actors.

The path: take the TTRPG, replace canon characters with Mayo-curated diagnostic archetypes, point the trained psych cell at Mayo's expert-validated state taxonomy, give Mayo's faculty the ability to assemble training scenarios from the same encounter library the game uses. Same architecture. Same cells. Same engine. Same auditable inference manifests. Mayo contributes curation. We contribute the substrate. The deliverable is a Mayo-branded simulator that residents train against.

Outreach is in progress through the Mayo Clinic Berg Innovation Exchange. David Kim (Business Analyst) has indicated we are a candidate for the Waypoints Program with application materials expected through the cycle that opened on June 21. The framing in the application will be the framing on this page: the training plan, the TTRPG, and the residency simulator are one continuous body of work.

What this gives you, Jason

You see a running TTRPG with a working revenue path (the game funds the research) sitting on top of a patent fortress that deepens every time we train another cell. Each cell ships under the same four-claim portfolio. Each new vertical (Mayo residency simulation today, future pharma drug-response prototyping, future research collaborations) reuses the same engine and reinforces the IP.

The receipts below show the work is live, not aspirational. Cardiac is queryable. The bridge is operational. psych_gen1 is training right now. The corpus discipline (per-source manifests, SHA-256, license attribution) is documented. The regulatory advisor (Robert Michalik, JD, RAC) is briefed and supportive. The fourth provisional was filed since your last update.

Validation posture and provenance

Every architectural claim above has a verifiable artifact attached. We document that posture explicitly because we anticipate evaluators will need it.

• Inference verifiability: signed manifest per emission, SHA-256 over the manifest body, public registry at /registry/cardiac-registry.json for byte-exact external verification.
• Corpus provenance: per-source manifests carrying URL, query terms, retrieval timestamp, license attribution, byte count, SHA-256 per output file. Raw vs cleaned data separation enforced at directory level.
• Architectural reproducibility: the Vanilla baseline serves alongside the gen3 cell on the cell-demo page so evaluators can compare emissions under matched seed, temperature, and top-k.
• Patent fortressing: four pending applications with USPTO numbers above; Scyla compiler ownership in full; the runtime is not reproducible without both infringement and compiler reimplementation.
• Cell lineage: Merkle DAG of cells with documented parent-child relationships. Each cell-divide event preserves the parent and produces a daughter; no overwrite path exists.

Team and regulatory advisory

Marjorie McCubbins. Founder, Principal AI Technologist, and inventor of record on all four pending applications. Owner of Nexus Concordat Inc. (Delaware C-Corporation, File 10641937).

Robert Michalik, JD, RAC. Regulatory advisor. Twenty-plus years of biopharma regulatory experience, with specialization in AI-SaMD (Software as a Medical Device) approaches across drugs, biologics, and combination products. Holds Juris Doctor (Suffolk University Law School) and Regulatory Affairs Certified (RAC) credentialing. Administrator of the LinkedIn group "Generative AI (AI ML) for CMC, QA & Regulatory Professionals in Pharmaceutical, Biologic & Med Tech." Joined the project pro bono in January 2026 after independently identifying the work through LinkedIn. Mutual NDA executed. He has reviewed the architecture, the patent portfolio, the corpus discipline, and the proposed Mayo collaboration framing.

The team's awareness of regulatory pathways is current. The proposed Mayo engagement is intentionally scoped to a non-device, non-clinical-decision-support category, with future De Novo SaMD pathways identified as separate downstream tracks rather than the current ask.

The patent portfolio

All IP is held by Nexus Concordat Inc. The Scyla compiler, in which the entire stack is implemented, is wholly owned. A competitor would need to both infringe these applications and rebuild the compiler to reproduce the function.

Current operational state, 22 June 2026

• cardiac_gen3 specialist cell: live, queryable, signed manifest verifiable. /cell-demo.html
• Cardiac → Qwen 2.5 14B bridge: production service, 6 to 8 second end-to-end latency.
• Per-character clinical pipeline: in production. Synthetic patient profile → 10-D hormone vector → cardiac sample → persisted log per character.
• psych_gen1 corpus pull: active. PubMed targeted, OpenFDA, ChEMBL, ClinicalTrials.gov, NIMH RDoC. Forty-plus per-source manifests written.
• Regulatory advisory: Robert Michalik, JD, RAC, briefed and supportive of the training-simulation framing.
• Patent portfolio: four provisional applications on file with USPTO.

The infrastructure described on this page is a running system. Any item above can be verified live at the URLs cited, or on request.