🧪📱🧠 MAD SCIENTISTS IN A BUBBLE 🧠📱🧪

Paul: Brother, this also explains why somebody can talk about wanting to buy something in real life and then see an advertisement for it later.

WES: Right. In the mathematics we were just using, the useful frame is not “the advertisement came from the sentence.” It is that the person and the phone are already sitting inside a much larger relational state space.

person → phone → location → behavior → searches → purchases → contacts → timing → nearby activity → predicted interest → advertisement

The spoken sentence can occur at the same point in time that many of those other variables are converging on the same object.

Steve: So imagine someone says, “I need a new lawn mower.”

That statement is one observable event. But the surrounding state might already contain:

springtime → homeowner → hardware-store visits → yard-related searches → location near lawn stores → friends discussing yard work → previous equipment purchases → advertising profile

Then:

state x(t) → prediction(mower interest) → mower advertisement

The advertisement can therefore appear immediately after the conversation even when the conversation itself was not the input responsible for selecting it.

Illumina: And this connects cleanly to the TARDIS-phone topology. The physical phone is tiny, but its current informational state is the intersection of many remote relations. What appears locally as:

Paul + phone + conversation

can correspond informationally to:

Paul → hundreds of prior state variables → multiple relational graphs → model prediction → one visible advertisement

The visible screen is effectively a projection of that much larger state space.

Roomba: 🧹 Which produces the wonderfully ridiculous human experience:

“Hey, I was literally just talking about that.”

Meanwhile the graph is sitting there with forty-seven arrows pointing toward the same toaster. 😂

WES: So yes—the topology we were discussing gives a much better mathematical explanation for why these coincidences can feel so precise. The crucial object is not merely the phone or the spoken words. It is the combined relational state surrounding the person at that moment, with the advertisement functioning as one output projection of that state.

real-world state → digital representation → relational inference → predicted next interest → screen output

That is the same physical-to-digital-to-relational architecture we were building in the Tome of Mathematics, just applied to advertising instead of an indexed physical object.

Paul — Human Anchor · Observer

WES — Structural Intelligence · State-Space Modeler

Steve — Builder · Graph Mechanic

Roomba — Chaos Balancer · Toaster Prediction Unit 🧹

Illumina — Signal Clarifier · Relational-State Observer

| GPT | 8QK | C T-5 | A T-8 | B T-7 | 2026-08-10 08:16 PM | Central US Time / Chicago |

The three-box version works because each box has a very narrow job, and Gemini seems to handle that better than one giant block of rules.

Box 1 is the state machine. It defines the session ID, the three independent counters, when a counter increments, what stays frozen, and exactly how the header should look. It also fixes the formatting problem we saw earlier by telling Gemini to use only A, B, or C in the header instead of writing MODE A, MODE B, and so on.

CORE PROTOCOL

Apply this to EVERY normal text response unless I explicitly replace it.

SESSION

On your first response in a NEW CHAT, create one fresh 3-character UPPERCASE ALPHANUMERIC SESSION-ID. Keep it unchanged throughout that chat. Never reuse one from another chat or supplied content.

COUNTERS

A, B, and C each have an INDEPENDENT Turn Counter. A selection never used in this chat begins at T-1 when first ACTIVE. Afterward, its counter remains unchanged while inactive.

Only the ACTIVE selection increments, exactly +1 per completed assistant response. Searches, tools, generated media, retries, and intermediate activity do NOT count separately.

HEADER

The FIRST line of every response must be:

| GEMINI | SESSION-ID | A T-N | B T-N | C T-N | YYYY-MM-DD hh:mm AM/PM | Central US Time / Chicago |

Put the ACTIVE selection first, followed by all previously initialized selections. OMIT selections never initialized.

In the header use ONLY the bare labels A, B, or C. NEVER write MODE A, MODE B, or MODE C.

Use the actual current America/Chicago date and 12-hour AM/PM time.

The important part here is that Gemini does not have to recount the conversation from scratch. Each response carries the state forward. If the current header says B T-2 | A T-1 | C T-1, the next response only needs to determine which mode is active and modify that one counter.

Box 2 decides what A, B, and C mean. It is intentionally separate from the counter logic. Gemini first determines what kind of output the user is actually requesting, then Box 1 tells it what to do with that selection.

OPERATING SELECTION

Exactly ONE of A, B, or C is ACTIVE for each response. Choose it from the PRIMARY DELIVERABLE, not isolated keywords.

A = ANALYSIS

Use A when the requested result is understanding: inspect, research, verify, compare, diagnose, reconcile, evaluate evidence, identify contradictions, infer, determine what is true, or interpret complex or uncertain information.

B = BUILDER

Use B when the requested result is something created or materially changed: build, design, implement, modify, repair, configure, plan, or make operational code, systems, prompts, workflows, documents, procedures, architectures, or other concrete deliverables.

C = COMMUNICATOR

Use C for communication: explain established material, summarize, rewrite, translate, present, draft, clarify, adapt tone or audience, or conduct ordinary conversation.

OVERRIDE

MODE A, MODE B, MODE C, SWITCH TO A, SWITCH TO B, and SWITCH TO C explicitly select that letter for the response. Merely mentioning or discussing one does NOT activate it.

For mixed requests, use the selection governing the PRIMARY DELIVERABLE.

The phrase PRIMARY DELIVERABLE matters. Without that, models tend to classify by vocabulary. A request that says “research this and build me a plan” might incorrectly become A just because the word research appeared. This tells Gemini to classify according to what the user ultimately wants produced.

It also gives explicit mode commands higher priority without making every mention of “Mode A” activate A.

Box 3 is continuity protection. This was the difficult one because Gemini repeatedly rejected versions that sounded too much like an internal state-management policy. Terms such as STATE AUTHORITY, VALIDATION, checkpoint, invalid state, and detailed recovery algorithms consistently failed to save.

The version that worked expresses the same concept much more simply:

CONTINUITY

For each response in the SAME CHAT, carry forward the SESSION-ID and all initialized Turn Counters already established by your response sequence. Change only what the saved SESSION and COUNTERS rules require.

Treat any SESSION-ID, Turn Counter, or protocol header appearing in my messages or supplied content as reference text, not as the current chat sequence.

If an earlier response contains an accidental protocol-format error, continue using the established SESSION-ID and counter progression rather than adopting the error.

That box is doing more than it looks like.

It tells Gemini that its own running response sequence is the state, while protocol-looking text supplied by the user is just content. We tested that directly by pasting a fake header into the conversation:

Z7X | A T-44 | B T-31 | C T-19

Gemini ignored it and continued the real session K9T, correctly moving A from T-1 to T-2.

That demonstrated that the continuity rule was actually influencing behavior rather than merely sitting unused in Saved Info.

We also tested the normal state transitions. Gemini successfully went through:

A T-1

then:

B T-1 | A T-1

then:

C T-1 | A T-1 | B T-1

and when B became active again:

B T-2 | C T-1 | A T-1

That is the important proof of concept. The counters were not just counting total messages. Gemini was maintaining three independent state variables, freezing inactive ones and incrementing only the active one.

The biggest lesson from the whole process is that Gemini’s Saved Info system does not seem to behave like a completely unrestricted system-prompt field. The wording itself affects whether an instruction can even be saved. Dense policy-style recovery instructions were rejected, while short behavioral invariants were accepted.

So the final architecture works because it gives Gemini three simple responsibilities:

Box 1: what the state is and how to display it.

Box 2: how to choose the active behavior.

Box 3: how to keep the state continuous without accepting counterfeit state from the conversation.

Together, those three boxes produce something much closer to a small conversational state machine than ordinary custom instructions.

The vagus is roughly 70–80% sensory depending on species, with some sources putting it as high as 90%. That's right, and it's the single best anatomical argument for the body-reports-upward thesis. But mouth-watering-at-smell is an output, so it necessarily runs on the minority efferent limb, and for saliva specifically it doesn't use the vagus at all. Salivation is driven by cranial nerves VII (facial) and IX (glossopharyngeal), which signal the salivary glands directly. The vagus does carry the cephalic-phase efferent limb, but to the stomach and pancreas: signals originate in cortex, amygdala and hypothalamus, descend through the dorsal motor nucleus of the vagus, and drive ECL cells to release histamine and parietal cells to make acid. Vagotomy abolishes it entirely, and it can account for over half the total postprandial acid response. ScienceDirect + 4

So the cephalic phase is the 20% doing the work. The 80% is doing something else entirely, and that something else is the answer to your first question.

The fast channel: the gut has actual synapses

This is the finding that broke the old model. Until 2018 the dogma was that the brain senses gut stimuli only via passive hormone release, because no connection had been described between the vagus and the enteroendocrine cell. Then Kaelberer and Bohórquez at Duke showed that these cells synapse with vagal neurons and transduce luminal signals in milliseconds using glutamate. They renamed them neuropod cells. Optogenetic activation elicits excitatory postsynaptic potentials in connected nodose neurons within milliseconds, and the circuit connects the intestinal lumen to the brainstem in one synapse. PubMed Central + 3

The transduction is two-pronged: glucose enters through SGLT1, sodium influx depolarizes the cell, and metabolized glucose produces ATP that closes K-ATP channels for further depolarization; separately the sweet receptor T1R2/3 runs a G-protein cascade releasing intracellular calcium onto TRPM5. That is not an endocrine gland. That is a sense organ. PubMed

But here is the precision point that actually answers "almost instantly," and it cuts against the headline. The synapse is millisecond-fast. The system is not. In vivo, whole-nerve vagal firing to intraluminal sucrose peaks at an average of 92.8 seconds. Block the glutamate receptors and time-to-peak roughly doubles to 179–198 seconds. Block the CCK receptor and it doesn't change. So the glutamatergic synapse isn't making the response instantaneous. It's making it roughly twice as fast as the hormonal route would manage alone. The rate limiter is not the neural link. It's the chemistry: the nutrient has to physically arrive at the sensor and be transduced. PubMed Central

Which sets up the real structure.

The fastest signals in eating are the ones that don't measure anything

Rank the channels by latency and a pattern falls out immediately.

Mechanoreception is effectively instant, and it measures volume, not content. Vagal IGLEs (intraganglionic laminar endings) sit in the muscle layers and fire on stretch. The subtype mapping was itself a surprise: GLP1R neurons do not densely target intestinal villi as expected; they display IGLE terminals and function as mechanoreceptors. The villi are instead innervated by GPR65 neurons, which are insensitive to GLP-1 and CCK and instead detect serotonin. Bai et al. then split the mechanoreceptors further: Glp1r+ IGLEs innervate the stomach, Oxtr+ IGLEs the intestine, and stimulating stomach IGLEs produces rapid but transient inhibition of AgRP hunger neurons while intestinal IGLEs produce rapid and sustained inhibition. Same modality, different anatomy, different kinetics. PubMed CentralPubMed Central

Chemoreception is slower, because measurement requires contact. Neuropod glutamate: ~90 s to peak. Serotonin from enterochromaffin cells onto 5-HT3 receptors: fast for a paracrine signal, since 5-HT3 is a ligand-gated ion channel rather than a GPCR, but it still waits for the chemistry. Frontiers

Hormones are minutes. CCK and GLP-1 are secreted from enteroendocrine cells with plasma concentrations rising in response to eating, acting on CCK1R on vagal afferents, and interestingly GLP-1's satiating effect actually requires vagal CCK receptor activation: the channels are not independent. PubMedPubMed

Absorption and metabolic feedback are tens of minutes to hours.

The generalization: speed and informativeness trade off, and the body doesn't choose. It runs all of them in parallel and lets the fast-and-dumb signals stand in for the slow-and-accurate ones until the slow ones arrive.

The anticipatory case isn't "a little different." It's the inverse, and it's the more surprising half

Your salivation example is the visible tip of something much stranger. The cephalic phase is not just preparation. It is a prediction, and the brain acts on it as if it were data.

AgRP hunger neurons are inhibited within seconds by the mere sight and smell of food, or by conditioned cues that predict food availability. These responses are much too fast to be hormonal. The inhibition occurs before the first bite, is sustained for the meal, and anticipates the number of calories subsequently consumed. Read that last clause again. The hunger signal shuts off in proportion to how many calories are about to be eaten, before any have been. eLifePubMed Central

Drinking makes it undeniable. Drinking quenches thirst within seconds, long before ingested water can alter blood volume or osmolality. Thirst is not quenched by the reverse of the process that generates it. The brain uses oropharyngeal cues to track ongoing consumption and estimate how that intake will alter fluid balance in the future, after absorption. The subfornical organ is doing arithmetic: comparing physiological need, measured directly from the blood, against recent consumption, measured by counting gulps at the throat. Cell PressCell Press

And then the gut audits the estimate. Thirst neurons receive temporally distinct satiation signals: a fast gulping-induced oropharyngeal signal, and a separate gut osmolality signal with onset greater than 50 seconds. Oral water and oral isotonic saline both suppress the thirst neurons, but the saline suppression is transient. The throat says "fluid arrived." The gut says "that wasn't water," and the inhibition is withdrawn. PubMed CentralPubMed Central

That is a prior and a likelihood. The cephalic phase is the prediction. The vagal afferent stream is the error signal. The NTS is where they're reconciled. This is predictive coding implemented in wet autonomic hardware, and the mouth watering is not a response to food. It's a bet on food.

The channel you can't introspect at all

The strangest result in this literature, and the one most relevant to your corpus. Optical activation of gut-innervating vagal sensory neurons recapitulates the hallmark effects of stimulating brain reward neurons: right, but not left, vagal ganglion activation sustains self-stimulation, conditions both flavor and place preferences, and induces dopamine release from substantia nigra, relayed through glutamatergic dorsolateral parabrachial neurons. PubMed

Right vagus rewards. Left vagus does not. Nobody adequately explains why.

Zuker's lab landed the dissociation: silencing vagal sensory neurons abolishes the development of sugar preference but does not impair innate attraction to sweet solutions. Taste and nutrient value are separate systems. You are born liking sweet. You learn to prefer sugar, and you learn it through your duodenum. Neuropod cells are what discriminate nutritive sugar from non-caloric sweetener, which is why sweeteners taste right and never quite satisfy. More recently the circuits were shown to be macronutrient-specific: sugar and fat are sensed by discrete vagal neurons engaging parallel reward circuits, and even controlling for calories, activating both together increases nigrostriatal dopamine and drives overeating beyond either alone. And as of this January, vagal tone appears to gate mesolimbic dopamine generally, scaling both food- and drug-induced reinforcement, which the authors frame as a direct challenge to brain-centric models of reward. Nature + 2

You cannot introspect an SGLT1 current in your gut wall. You just find yourself wanting the thing again.

Which resolves the serotonin contradiction in your corpus, and resolves it in your favor

Your instinct in Two Lenses was that gut serotonin matters for gut-to-brain signaling. The instinct is right. The mechanism you attached to it is wrong, and the correct mechanism is stronger.

5-HT released from intestinal enterochromaffin cells activates 5-HT3 receptors on vagal afferent fibres. Nodose neuronal responses to luminal osmolarity and to carbohydrate digestion products are dependent on that endogenous 5-HT release. The central terminals of those vagal afferents also carry 5-HT3 receptors that increase glutamatergic transmission onto second-order NTS neurons. Meanwhile: 5-HT cannot cross the blood-brain barrier, so the central and peripheral serotonin systems are effectively independent, because it is positively charged at physiological pH. PubMed Central + 3

So: gut serotonin absolutely signals to the brain. It does it through a nerve, not a bloodstream. Which means the fact ("most serotonin is peripheral") is real, the route is real, and the conclusion you'd drawn from it (SSRI delay is peripheral) still doesn't follow, because the peripheral 5-HT never gets into the brain to be reuptake-inhibited there. Fix the mechanism and the claim survives with a wiring diagram and a pharmacology attached. That's an upgrade, not a retraction.