Live demo · 01 June 2026

The cell speaks.

Type a clinical phrase. The cardiac-gen3 specialist (trained through the 34-head Hodgkin-Huxley cascade, finished 01:55 UTC this morning) generates 32 tokens autoregressively. You get the decoded sentence, the signed inference manifest, the verify result, and the raw JSON. Real inference. Real signature. Verifiable by SHA-256 against the public registry.

New · cardiac → 14B bridge live
Every inference on this page now also fires the patch: the cardiac cell's raw substrate output is handed to Qwen 2.5 14B as the biological prior, and the 14B translates it into clinical English. Cell speaks biology. 14B speaks language. The bridge is the conductance layer. Architecture: jason-plan.html · cell as opcode tree.
Queued · psych_gen1, sister cell in training
cardiac_gen3 is the first specialist. psych_gen1 is the second, in queue: same architecture, psychiatric corpus, four-brain ablation plus vanilla baseline. The proof the architecture generalizes.

Query the cell

Type clinical words this cell knows. One query per 10 seconds per IP. Each query runs ~3 to 8 seconds.

Try: heart · chest pain · patient stemi · warfarin metoprolol · hypertension · cardiac surgery · aortic stenosis. Unknown words come back with an error listing which ones the cell does not know.
Our cell said cardiac_gen3_bio · chemistry-conditioned
Vanilla transformer said same data · no cascade
The patch · cardiac substrate translated by Qwen 2.5 14B cardiac → 14b bridge
The cardiac cell speaks in biology. The 14B speaks in language. The bridge is the conductance layer between them. Architectural plan: jason-plan.html.

What you are actually querying · both columns

Every query you run on this page fires two HTTP requests in parallel: one against our chemistry-conditioned cardiac cell, and one against a standard textbook transformer trained on the identical corpus. Same prompt, same seed, same temperature, same top-k. The point is the side-by-side. Both models in detail.

Left column · our cell

Cell ID (registry): 57cb59a36c665bd75ab38f7927c61c60d9c64cf367c6b2465e26b8a2a1603d01 · kind: cardiac_gen3_bio · generation: 3 · parent: cardiac_gen2 (Merkle DAG live at /registry/cardiac-registry.json).

The model is the chemistry-conditioned twin of the voice-only gen3. Every training step ran through the 34-head 4-tier Hodgkin-Huxley cascade (8 amino acid + 6 precursor + 8 neurotransmitter + 12 brain region conductance heads, frozen at cascade_lr=0). The cascade conditioned the voice's attention during training. The voice tower is what you sample at inference. Per-request runtime cascade weights are re-initialized fresh, so the live cell_id in the manifest will differ from the registry's; the opcode-tree hash and signature still verify byte-exact internally. Persisted-cascade builds are queued for the next pass.

Trained on: 20,056,125 tokens of cardiac and clinical-trial substrate over 600,000 effective steps (5K → 15K → 35K → 76,325 vocab via the cell-divide primitive). Carrying six architectural handicaps vanilla does not: visit-cap K_max=6 (each token gets at most six training visits before it freezes), batch size 1 (one token at a time, 64× less gradient signal per step than vanilla), three forced cell-divide weight inheritances mid-training, cascade forward pass at every step, deterministic corpus walk (no random batching), and character-level vocabulary (every long drug name is one row, no BPE).

Right column · vanilla baseline

Architecture: single-block PyTorch transformer. Vocab: 76,325 · d_model: 256 · n_heads: 8 · d_ff: 512 · seq_len: 16 · params: 39.7M (within an order of magnitude of our 32M cell). Pre-norm, GeLU FFN, causal multi-head attention. No Hodgkin-Huxley gates. No nt_proj. No entity embedding. No ctx_proj. No cascade. No signed manifest path. No cell-divide. No Merkle lineage. Plain textbook transformer.

Trained on: the identical cardiac_mega_tokens.txt (20,056,125 tokens) for 300,000 AdamW steps at batch size 64 (= 19.2 million training examples, vs our cell's 600K single-token steps). Final loss 4.76. Sample latency ~0.8 s per query through the public domain.

Why we built it: the apples-to-apples we owed Jason. Same data, same vocab, same architectural scale, no architectural constraints — what does standard get? Surface fluency, immediately. Coherent clinical-trial sentences. No verifiability. No specialty isolation. No lineage. The vanilla side of this page is the honest counterfactual.

What this is telling us: the surface fluency gap between our cell and vanilla is a function of the visit-cap and batch-size constraints — both of which are architectural choices we made for verifiability, not for performance. Drop them and the architecture matches vanilla on fluency. Keep them and we ship the verifiability vanilla structurally cannot produce. The fact that our research substrate, carrying all six handicaps, lands in the same cardiac vocabulary distribution as the unconstrained textbook architecture is itself the headline.

The verification chain

Every inference response contains a signed manifest binding the sampled token to the cell ID, the opcode-tree hash, the sample parameters, and a SHA-256 signature over the manifest body. The verify field is the server's check (recomputes opcode-tree hash and signature, compares byte-exact). An external party can do the same: fetch the public registry JSON, recompute SHA-256 over the opcode tree of the cell ID in the manifest, and confirm the opcode-tree hash matches. The architecture is public; the weights are sovereign.

The registry is at /registry/cardiac-registry.json. The opcodes table is at /opcodes.html. The architectural plan is at /jason-plan.html.