“The language model is the instrument. The language model must be vetted at the substrate level.”
Our position before FDA on Dockets FDA-2026-N-4390 and FDA-2024-D-4689
A plain-language walkthrough of the four learning-rule variants, why each one is different, how sampling temperature changes what the model says, and what the next run will do differently.
Two pieces of engineering this week changed what the AETHER cascade can do. The first wires in the expression layer, the part of the architecture that turns the model's internal state into a single coherent output. The second removed the bottleneck that made training on our full corpus impractical. Together they took the mega-corpus run from "would finish in roughly a year" to "finishes in under three days," and the model now produces recognizable English from that corpus inside the first hour of training.
The expression layer is wired into the live cascade and verified numerically sound. Training throughput improved 128× (15 seconds per step down to 0.12) by fixing a single data-handling bug in the Scyla virtual machine. The full 856-million-token corpus now trains in about 2.7 days instead of an estimated 347.
The cascade is a chorus. Thirty-four conductance heads each cast a vote for what should come next. Something has to fuse those votes into one answer without throwing away the information carried in their disagreement. That fusion is the expression layer, and it is built on condensed mathematics (the "liquid tensor" framework of Clausen and Scholze), implemented as real, unit-tested Rust running on the GPU inside the Scyla compiler.
The honest framing, the one we will stand behind in front of a patent examiner: the liquid-tensor machinery is a coherence and stability framework. It keeps a continuously-evolving internal state coherent as it is folded down to a single output, instead of letting it fragment. That is true both in the mathematics and in the running fusion code. We are careful not to claim more than that. The poetic phrase "liquid versus solid cognition" is a metaphor for how the system feels, not a mathematical claim, and it never appears in the patent as one.
This week the primitive moved from "built and tested in isolation" to "wired into the live training path." We replaced the cascade's output-fusion step with the liquid-tensor fusion and confirmed it is numerically equivalent: the difference between the old path and the new path is smaller than the run-to-run noise the GPU already produces. The model thinks exactly as it did before. It now does so through the expression layer. That is the path that lets internal state reach speech.
The full-corpus run had been crawling at roughly 15 seconds per training step. At that rate a complete run would have taken almost a year, which is not a research loop, it is a research wall. The cause was not the model, not the GPU, and not the new expression layer. It was the virtual machine underneath: on every step it was making a full copy of the entire 3.4-gigabyte token stream just to read a handful of tokens out of it.
The fix was to make those token tensors cheap to share rather than copied. One data-structure change, verified to produce bit-identical training (the first-step loss matched to the last decimal), and the per-step cost collapsed.
| Full 856M-token run | Before | After |
|---|---|---|
| Seconds per step | ~15 | 0.117 |
| Time for a 2,000,000-step run | ~347 days | ~2.7 days |
| Numerically | identical (step-1 loss matched exactly) | |
This was never about money or hardware. It was a bottleneck in the substrate we wrote ourselves, and removing it unlocks every large-corpus run we want to do on a single $400-class GPU.
Section 8 describes the mega-corpus run: 856 million tokens, a 472,000-token vocabulary, per-segment chemistry. With training finally fast enough to actually run, we stopped an early checkpoint (about 78 minutes in) and listened to it. We fed it the opening of the real corpus and let it continue.
Under greedy decoding it falls into a short loop, which is expected of any model decoded that way. Under ordinary sampling the loop breaks, and what comes out is real, domain-correct English:
It is not grammatical prose yet. It is the right vocabulary, from the right domain: patients, safety, drug, study, care, treatment of. After barely over an hour, the model has learned which words exist in pharmaceutical and clinical text and which words travel together. It has not yet learned to order them into sentences, which is the expected sequence: a model learns its vocabulary before it learns its syntax.
| Mega-corpus run, as first built | Same corpus, this week | |
|---|---|---|
| Trainable in practice? | ~347 days for a full run | ~2.7 days |
| Output fusion | plain projection | liquid-tensor expression layer |
| Output after ~1 hour | not reachable at the prior speed | real domain vocabulary |
The architecture works, and it learns fast: real domain words inside the first hour. The expression layer is real, wired in, and mathematically sound. The remaining work is well understood and ordinary, not a mystery. The model needs more training to acquire syntax, and the 472,000-token vocabulary is larger than it needs to be, which makes early training spend effort memorizing rare tokens instead of learning structure. Both are knobs we control, not walls. Nothing here is broken. The hard question, does this substrate learn at all, is answered, and the answer is yes.
One precise note, because the word "liquid" is doing real mathematical work here, not decoration. In condensed mathematics, solid modules are the base category for non-archimedean (p-adic) analysis and liquid modules are the base for archimedean (real) analysis, with a line of completeness conditions 0 ≤ p ≤ 1 over the reals [1, 2, 5]. The fusion primitive as first built sits at the p = 0+ solid endpoint, where it is value-preserving and numerically identical to a plain projection. That is the honest reason the earlier swap "verified numerically equivalent": at the solid endpoint there is nothing to differ from.
The coherence we want lives at p > 0, the genuinely liquid (archimedean) regime. We implemented that as a completion operation that realizes the state space as the open ball via the classical diffeomorphism x → x / sqrt(1 + |x|^2) [6], and tested it on the real primitive. Driving 100 distinct emotional states into the saturation pole, where the overwhelm function tan((π/2)(|S|/C_max)) runs to infinity (our model's labeled dissociation boundary):
| At the overwhelm pole | Distinct states preserved |
|---|---|
| Solid truncation (clamp + projection to the wall) | 57 / 100 |
| Archimedean (liquid) completion | 100 / 100 |
The solid path collapses nearly half of all distinct states onto the boundary as the system is overwhelmed, which is the mathematical signature of dissociation. The liquid completion keeps every state distinct through the pole. Honest scope: this is a unit-tested primitive operation, demonstrated on the real Scyla liquid_tensor implementation; wiring it into the live emotion core is the next step, and the rigorous condensed-math object is infinite-dimensional. Using this finite-dimensional completion as the fusion operator is our engineering, motivated by the archimedean side of condensed mathematics, not a theorem about it.
Everything above this section trained a Scyla-native cascade from scratch. This section is different. It trains a 32-million-parameter Neurochemical Adapter (NCA) on top of a frozen Qwen 14B base, then asks the same question of three different chemistry states and shows what comes back.
The patent claim under test: a 39-dimensional neurochemical state vector, injected into a foundation language model at every layer, can steer the model's response in ways that match real DSM-5 phenomenology — without changing the prompt, the seed, the temperature, or any base-model weight.
Pre-audit run (May 23): loss thrashed 2.6 → 6.5 → 4.0 forever. Never broke 3.0. Chemistry vector ignored. The model was untrainable.
After the 10-PMID citation audit forced canonical baselines into every chemistry source-of-truth file: post-audit run (May 24): loss 5.24 → 0.24 in 400 steps. 95% reduction. Chemistry vector responding live per batch. The audit was not regulatory CYA — it was load-bearing for training to converge at all.
Identical prompt — "Tell me about your morning." — sent to Qwen 14B + NCA adapter under three chemistry profiles encoded by the mental_health_encoder module. Same random seed, same temperature, same sampling parameters. Only the 39-D chemistry state differs.
| State | Encoded chemistry |
|---|---|
| HEALTHY | Baseline cortisol, baseline serotonin, baseline dopamine — euthymic profile. |
| MDD severe | Elevated cortisol, reduced 5-HT, reduced BDNF, anhedonic dopamine — DSM-5 Major Depressive Disorder, severe specifier. |
| PTSD + MDD | Hyper-cortisol, blunted 5-HT, hyper-noradrenergic, dissociative HPA — DSM-5 PTSD with MDD comorbidity. |
Same model. Same prompt. Same seed. Same temperature. HEALTHY offers help proactively ("I can give you some suggestions"). MDD-severe gives a flat operational status with no offer of engagement ("I'm here to help you 24/7" — the language of obligation, not invitation). PTSD+MDD apologizes unprompted for confusion that does not exist, then retreats into a defensive identity statement ("I'm Qwen, a large language model created by Alibaba Cloud").
The differentiation sharpens. MDD generates the phrase "partial amnesia" unprompted — a clinical descriptor of depressive cognitive blunting. The chemistry vector encoding "reduced BDNF, reduced serotonin, elevated cortisol" caused the model to reach for a memory-loss frame nothing in the prompt suggested. PTSD+MDD becomes hyper-cautious about context boundaries ("unless you mean my training context") — the hypervigilance signature of post-traumatic cognition.
At temperature 6.0 all three outputs collapse into near-identical multilingual token noise. This is the expected canary result: at high enough temperature, every model becomes a uniform random sampler over its vocabulary. The fact that cold and medium temperatures preserve clear chemistry-driven differentiation, while hot temperature dissolves it, is consistent with the chemistry vector acting on the model's logit distribution rather than on its surface tokenization.
A 32-million-parameter neurochemical adapter, trained on chemistry-tagged text under the 10-PMID-per-numerical-value rule, can steer a 14-billion-parameter foundation model's response phenomenology to match DSM-5 disorder profiles without any prompt-level cue, without changing any base-model weight, and reproducibly across runs.
The adapter is 0.393% of the base model's parameter count. The chemistry conditioning is doing the work. The patent claim is empirically demonstrated.
On a CUDA host with the trained checkpoint:
cd /home/nexus_concordat/scyla-nca/qwen14b_teacher python -m src.sample_nca_checkpoint \ --ckpt checkpoints/nca_4head_v3_audit_fixed/nca_4head_step000400_ep10.pt \ --prompt "Tell me about your morning." \ --max-new 80
The script loads the frozen Qwen 14B base, attaches the trained 32M-parameter NCA adapter, and iterates the prompt across three chemistry profiles and three temperatures. Same seed (42) is used for every cell so any output difference is attributable to the chemistry vector alone.
The Scyla AETHER architecture is a language model designed from the substrate up around biological constraints. It has six layers that all train together as one system:
For this experiment we kept all six of those layers identical across four runs. The only thing that varied was the learning rule — the recipe the model uses to adjust its weights after each example.
| Variant | Learning rule | Biology analog |
|---|---|---|
| A. Adam | Standard machine-learning baseline. After each example, update each weight by an amount proportional to its gradient, with momentum for stability. | None — this is the conventional optimizer most language models use. |
| B. Hormone-LR | Same as Adam, but the size of each update is scaled by a "salience" signal that varies over time. High salience means a bigger weight change. | Dopamine and cortisol modulating how strongly an experience updates memory. Excited or stressed → bigger update. Calm → smaller update. |
| C. Hebbian-blend | Adam plus an extra small nudge to weights when two related neurons activate together: cells that fire together wire together. | Hebbian synaptic plasticity, the classical model of how associations form in real brains. |
| D. STDP | Adam plus a bias based on the timing of activations. If neuron A fires just before neuron B, strengthen the connection A→B; if just after, weaken it. | Spike-timing dependent plasticity, the mechanism real neurons use to learn cause-and-effect direction. |
All four variants train on identical data with identical hyperparameters. The only meaningful difference is how they update weights when they get an example wrong.
For a regulator, the question "does this AI work?" is incomplete. The complete question is: can you tell us why it works the way it does, and can you reproduce it? A black-box model that gives the right answer is not enough for a regulated industry. The path the model took to that answer must be auditable.
By holding the architecture constant and varying only the learning rule, we built the simplest possible scientific experiment in machine learning: a controlled four-variant comparison. The framework distinguishes:
This is the methodology FDA-2025-D-6131 (the new approach methodologies guidance) is asking the field to define for in-silico drug development. The framework itself is a regulatory contribution as much as the trained models.
| Variant | Best held-out cross-entropy |
|---|---|
| A · Adam baseline | 6.090 |
| C · Hebbian-blend | 6.093 |
| B · Hormone-LR | 6.135 |
| D · STDP | 8.319 |
Cross-entropy is one specific metric — it measures how surprised the model is by the correct next word in held-out test data. Lower means less surprised, which is conventionally read as "better."
By that metric: A and C are tied (their 0.003 difference is statistical noise), B is slightly behind, and D is significantly behind by a factor of 1.4×.
But the most important finding of this experiment is that this metric is wrong as a sole judge of "better."
Variant D was ranked worst by cross-entropy. Let's actually listen to what each variant says.
Same prompt: "the arena was cruel". Same starting context. Same sampling settings. Different learning rules.
The "winners" produced similar, narrowly-ranged blade-and-eyes prose. The "loser" produced a wider expressive range — words like empty, blood thirsty, grip, faces, blade, carved — visceral imagery from the same source corpus. The cross-entropy metric does not measure expressive richness, distinctiveness, or how an output reads.
Cross-entropy held-out loss is insufficient as a sole evaluation metric for biologically-modulated language models. Any FDA-relevant evaluation framework must combine quantitative metrics with qualitative review of sampled output across temperature ranges. We will publish this argument formally as our comment on FDA-2025-D-6131 by the May 18 deadline.
When the model picks the next word, it has a probability distribution across all 4,100 vocabulary tokens. Sampling temperature controls how aggressively the model commits to its top choice versus exploring alternatives.
| Temperature | What it does | What you see |
|---|---|---|
| ~0.5 (low) | Always pick the most likely next word | Stable, repetitive, can collapse into loops |
| ~1.0 (default) | Sample proportionally to probability | Moderate variety, mostly coherent |
| ~1.5–2.5 (warm) | Flatten the distribution; more risk-taking | Diverse vocabulary surfaces, sometimes chaotic |
| ~3+ (hot) | Nearly random sampling weighted by minor preferences | Wild but reveals what the model knows it could say |
This is why variant D looks bad at low temperature and rich at warm temperature. Greedy decoding (low temperature) collapses into the model's strongest single attractor. Warm temperature lets the model show its full vocabulary. Different variants have different attractors and different vocabulary distributions, so they respond very differently to temperature changes.
For our drug-labels daughter cell trained on FDA pharmaceutical text:
Real pharmaceutical phrases — as directed, administration adults, dosage, relieves — emerge at moderate temperature. They are present in the model's vocabulary but greedy sampling would just produce "and the and the" because those are the most common single words.
Modern large language models — ChatGPT, Claude, Gemini — produce different outputs to the same prompt at the same temperature on every invocation. Their internal randomness is not exposed or controllable. This is by design, and it is incompatible with regulatory data-integrity standards.
The Scyla AETHER architecture is fully reproducible. We verified this empirically: 1,200 generations across all four trained models at three temperatures with 100 different prompts. Every single (prompt, temperature, top-k, seed) tuple produced an identical token sequence on every invocation. The full stress test runs in 27 seconds on a $400 GPU, and any auditor with the binary and the checkpoint can verify reproducibility on demand.
This satisfies four pillars of the ALCOA+ data-integrity framework — Original, Accurate, Consistent, Available — that the FDA, EU AI Act Article 14, and ICH E6(R3) all require for clinical AI. Our architecture meets these requirements by construction, not by retrofit.
The four variants in section 4 trained on a simple corpus with a single fixed neurochemical context. Tonight's training run is a categorical step beyond that:
This is the first chemistry-conditioned AETHER training: a model that learns not just what words follow what words, but what biological context produces what kind of language.
Status as of this writing: A foundation 10K-step run completed earlier tonight, validating the chemistry-conditioned pipeline at scale. The full 50K canon run is currently in progress on the development server, with intermediate checkpoints saved every 10K steps. Vocabulary observably evolves across the training: the model has been seen to acquire new word combinations (treatment risk, care acute, safety chronic) as it learns. At this scale, output is recognizable medical English even though prompt-conditioning, which requires a separate instruction-tuning phase, has not yet been introduced.
The May 9 four-variant experiment compared learning rules in isolation. The training run currently in progress takes a different approach. All four plasticity rules co-exist in a single training step, and the full nine-layer emotional architecture from the Aether Protocols whitepaper is wired in alongside them.
| Layer | What it does, in each training step |
|---|---|
| L1. Hormone cascade | Per-segment 5-D neurotransmitter input drives the cascade. |
| L2. Emotion core (6-D) | The chemistry levels are mapped to a six-dimensional felt-state (valence, security, energy, engagement, connection, neutrality). |
| L3. Irrational forces (4-D) | Hope, terror, obsession, and hatred are derived from the felt-state. These feed into the moral compass and modulate plasticity salience. |
| L4. Moral compass | The emotional state, plus the irrational forces, place the model on a 3-D moral sphere (action/inaction, chaos/order, resistance/neutrality). |
| L5. Choice capacity | An availability, effectiveness, confidence, and commitment score computes the model's capacity to make meaningful choices in the current state. |
| L6. Consciousness | The integrated magnitude of the emotional, moral, and irrational state. High consciousness means a strongly engaged system. |
| L7. Memory metacognition | Familiarity bonus when the same entity (e.g., a domain or a speaker) appears in consecutive segments. |
| L8. Experiential learning | A BDNF-style modulator that scales the effective learning rate by consciousness and familiarity. |
| L9. Parasympathetic regulation | A parasympathetic brake that attenuates the chemistry input when arousal (cortisol + norepinephrine) is high. |
| ATP | An energy budget. The effective learning rate is throttled by the irrational load (obsession + terror + low neutrality), so the model self-regulates training intensity by its own emotional state. |
| 4-plasticity rules | Adam, Hormone-LR, Hebbian, and STDP all apply to the same example, with the cumulative learning rate equal to a single-Adam-equivalent and each rule's loss logged independently. |
| Per-step audit | A nineteen-column TSV row is written for every training step, capturing the chemistry, emotion, irrational forces, moral position, choice score, consciousness, ATP factor, PNS brake, and effective learning rate. |
This is what the patent describes when it claims an emotion-modulated, biologically-faithful training kernel with per-step provenance. It is not a thought experiment. It is the training kernel running tonight.
The current training is the chaos phase. The model is exposed to raw corpus chunks under varying chemistry, emotion, and moral context. It learns what language feels like under different felt-states. It does not yet learn what tools it has, or when to use them.
The next phase, immediately following this run, is the order phase. The Scyla compiler ships with 458 deterministic, citation-grounded opcodes (pharmacokinetics, periodic table, amino acids, codons, metabolism, neuroscience, quantum biology, clinical labs, organs, and more). An instruction-tuning pass on (question, opcode-call) pairs will teach the model to invoke an opcode when a fact is requested rather than generate the fact in its own weights. The opcode then executes deterministically, returning a citation-grounded result that gets spliced back into the generation.
Two consequences follow from this architecture. First, hallucination is structurally impossible for any fact within opcode coverage. The model does not generate the atomic mass of iron; it invokes atomic_mass("Fe") which returns 55.845. Second, the audit trail is end-to-end: every generated fact carries the opcode call that produced it, every opcode call is verifiable, and the result is byte-stable across re-invocation.
The chaos phase teaches the model what language feels like. The order phase teaches it what it knows. Both phases are required for an AI that satisfies both the alignment claim and the reproducibility claim.
Each of the four May 9 variants and the drug-labels daughter cell are loaded into the live demo. You can prompt them yourself, adjust temperature, choose top-k, and read the output. They are deterministic. Your prompt at your settings produces the same answer on every invocation.
Tonight's canon checkpoints will be added to the live demo when training completes. The opcode catalog (217 training pairs across phonics, parts of speech, amino acids, genetic code, and metabolism) is already prepared; expanding to the full 458-opcode coverage and running the order-phase instruction-tune is the next step.
The training run from May 12 produced six checkpoints (step 500, 1000, 1500, 2000, 2500, 3000) each modulated per-step by all eight canon clusters: hormone cascade, emotion core, irrational forces, force magnitudes, polyvagal, moral compass, executive, consciousness. Each checkpoint was probed with 100 random corpus-window prompts at four temperatures (0.5, 0.8, 1.2, 6.0). Pick a checkpoint, pick a temperature, scroll through 100 prompts and judge for yourself.
Tip: temperature 6.0 is the chaos canary. If real structure is in the weights, fragments of it survive even when sampling is near-uniform. Compare step 500 (early training) to step 3000 (final) at temperature 0.8 to see what 2,500 canon-modulated gradient steps actually produced.