There Is No Pure Emotion: An Open Emotional Router for AI Agents, and the Field Model Under It

The Fort That Holds · Jimmy Thornburg · July 19, 2026

Abstract

Most systems that reason about emotion flatten it to a label — sad, happy, angry — and act on the label. This paper describes an alternative built for AI agents rather than for people to read: an open-source emotional router that teaches a foreign agent to resolve a moment of feeling into a braid — a dominant family, the feelings held under it, the body it lives in, the nervous-system register it runs in, and the humor that can carry it without cheapening it. The router ships thirteen emotion families not as atoms but as a weighting basis, and teaches the agent to assemble the combination live rather than pre-naming every one. Under the router sits a small design model, called the field: there is no such thing as a pure emotion; opposing feelings can be held in superposition without cancelling; feelings stack in a causal lock (control over rage over fear over an original wound); “pure” is the artifact of not looking closely enough; and humor is the load-bearing spine that both holds a dense feeling and diagnoses it. The colors, bodies, and nervous-system states are anchored in one person's own expressive systems (SongCraft's ChromaScribe color map and Emotion-Body Map), with cross-cultural affect research cited only as corroboration. We are explicit about what this is and is not: it is one person's map, not the map — a design instrument, not an empirical theory, and not a diagnostic tool. Internal agent-to-agent trials are reported as engineering dry-runs, not evidence, because an author grading work produced inside their own session is not an independent test. The single genuinely independent result we have — a different vendor's model, reading the repository cold over a connector — is reported in full as an anecdote, n = 1. The contribution is a working, inspectable artifact and a coherent lens, offered to be re-tuned, not a validated instrument.

1. The Core Claim

The claim the whole thing turns on is one sentence: there is no such thing as a pure emotion. You never actually feel “anger.” You feel anger-braced-over-fear, or anger-under-grief, or anger cut with a dry joke to keep it in the room. When a person, or a model, reports a single clean emotion, that is not the feeling — it is what the feeling looks like when nobody looks closely. The router's job is to stop the agent at the label and make it look closely: to route the field of the moment, never the note.

This is a design position, not a discovery. It is not offered as neuroscience or as a falsifiable theory of affect. It is a lens — a way of organizing an agent's attention so that its emotional output stops collapsing to one dimension — and its warrant is craft: it is drawn from years of one person's work building emotionally precise song, prose, and voice systems, and it earns its place by making agent output visibly less flat, not by a controlled measurement (see Sections 15–17 on exactly what evidence exists and what does not).

2. Feeling Flattened to Labels

The default failure mode is legible in almost any emotionally-aware system: a classifier tags a message sadness: 0.8; a character model plays “he is angry”; a writing assistant delivers grief as operatic weeping. Each is a collapse — a rich, layered state reduced to its single loudest component, then performed at volume. The reader feels the difference immediately, even when they can't name it: the difference between “grief, suppressed” and “grief, with the old anger still under it, held in a clenched body because someone else is in the room.”

The router treats the flattening itself as the defect to engineer against. Its first instruction is not “detect the emotion” but “look truly — do not collapse”: a beat that reads flat, bright, controlled, or “fine” is a signal of density held under a surface, not of a simple state. The most dangerous inputs are the ones that look easy.

3. Families as a Weighting Basis

The router ships thirteen emotion families (Appendix A). The critical design choice is that they are not a taxonomy of discrete emotions but a weighting basis — a set of dimensions over which any real feeling is a weighted combination. Routing a beat means setting weights across the whole set, then reading which dimension dominates and which two or three color it.

This is the same move a voice-engine makes on the register axis. In the house's writing system, a single deep authorial identity produces every pen by weighting a library of registers — one dominant, a small number of modulators — rather than switching between discrete authors. The field applies that structure to emotion: the “nine pure emotions” a naive system would enumerate are, in this model, simply the basis vectors seen one at a time, each one already a bundle weighted toward a single lean. The families are where you start assembling, not a menu you pick from.

Because the combinations are assembled live, the router cannot and does not pre-name them. It ships the basis, the law for combining it (Section 9), and worked examples of the assembly — not a lookup table of every braid. This is deliberate: a closed taxonomy would invite the agent to optimize the label instead of the feeling.

4. Non-Collapsing Superposition

The most useful borrowed intuition is from quantum superposition, with one deliberate inversion. In the pop image of Schrödinger's cat, the system is “both until observed,” and observation collapses it to one state. The field says the opposite about feeling: it is observation done carelessly that collapses a feeling to one pure state, and observation done truly that keeps the superposition intact.

Concretely, opposing feelings can be held at full charge at the same time without averaging to a muddy midpoint. Relief and grief at a death you also needed to be over; joy and terror in the same breath at a thing you want too much; love and dread pointed at the same person. The wrong move is to resolve these to “bittersweet” — the average, the collapse. The right move is to hold both at maximum and let the density itself be the content. The router names this in its output contract: certain pairs are to be held at full opposing charge and explicitly not resolved.

5. The Causal Stack

Feelings are not only braided sideways; they stack downward, each layer a defense against the one beneath it. The canonical stack the router routes from is:

control ← rage ← fear ← the wound

A sharp, over-controlled surface is usually a lock held over unbridled rage; the rage is frequently a more wearable face for fear; the fear traces to some original wound the person organized themselves around. “Control,” in this reading, is not one note — it is the whole column, compressed. The router's instruction is to trace the stack rather than stop at its top: name the lock, then what it locks down, all the way to the wound — while rendering only what the beat can carry. Rungs are optional; not every stack has a rage layer, and inventing one is its own kind of flattening. But the discipline is to look down the column, because the top of the stack is precisely the part designed to be seen.

6. Sharpness and the Collapse Artifact

Two consequences follow, and they are the model's sharpest working tools.

Sharpness is opposing weight held under a fragile lock. What reads as a “sharp” person or a sharp line is not a single intense emotion; it is a large opposing charge held under a tight surface — the density is the sharpness. This can present cold or bright: a too-even voice and a too-hard laugh are both locks. Reading sharpness as one loud note is the error; reading it as a dense field under tension is the route.

“Pure emotion” is a collapse artifact. When a beat genuinely appears to be one clean feeling, the model treats that as evidence of shallow looking, not of simplicity — a prompt to look again for the stack and the superposition, not a green light to play one note. The one case that most often fools a flattening system is numbness: a flat, “fine,” gone-quiet surface reads as calm but is almost never calm (Section 11 gives the physiological reason).

7. Humor as the Spine

Humor is not a garnish in this model; it is the spine — both load-bearing and diagnostic — and it does not work without the emotion under it. The router carries a compact account of why a joke lands, adapted from the house's HumorCraft work, as a rough product:

H = V × B(f) × ΔS

A joke's force is a violation (something is wrong, transgressed, incongruous) times a benign audit B(f) (the container is safe enough that the violation reads as play, not threat) times a script-shift ΔS (the sudden reframe from one reading to another). The benign-audit term is the load-bearing one for emotional work: when a wound is too hot for its container, B(f) approaches zero, the joke fails the audit, and silence is the honest move. This is why the same line — a wake-joke among people who share the grief — heals in one room and lands as a slap in another.

Humor is also where the two collapses reconcile. A punchline does collapse something: it resolves a small local prediction-superposition (the setup could mean A or B; the turn snaps it to B). But that small, chosen collapse is spent precisely so the large field-superposition — the dense grief-and-joy of the moment — can keep being carried instead of putting the person on the floor. The laugh at a funeral is the spine folding on purpose: a controlled small collapse that protects the big one. Humor is the load-bearing fold, and its failure to fire is itself a reading of how hot the field is.

8. A Context-Pack for Foreign Agents

The router is not a library you call and not prose you paste into your own product. It is a context-pack a foreign agent reads to learn a skill it does not have. You do not run it; your agent walks it. It imposes no voice: it teaches the move — resolve the moment into a braid — and the agent makes the braid in its own words, on its own surface.

Everything in the pack is wrapped in a Containment Protocol header stating that reading a family map means “you may route through this node,” not “become this voice.” This matters when the consumer is an arbitrary third-party agent: the material is information for routing, not an instruction to adopt a persona. The intended consumers are systems that must feel accurately — support agents, game characters, writing assistants, journaling companions, therapy-adjacent tools — and the router's promise to them is narrow and honest: it improves how they route feeling; it does not license them to tell a real person what they feel.

9. The Family Map and the Braid Law

Each family is a filled-in Spine — a node in the house's universal identity schema (the Fort Kit Spine), the same shape every voice and collaborator in the Fort uses, so the router is a graph of graphs. A family map carries: when to route here; the “4 a.m. truth” (the felt core under the socially-presentable version); the body it lives in; its sub-shades; its humor partner; the families it braids with (as edges, glossed from both sides); what to refuse; and how it assembles when dominant versus when a modulator.

The combining rule — the braid law — is: one dominant family, at most two modulators, plus the causal stack beneath and any superposition held across it. The cap is deliberate. It forces the agent to decide what leads instead of listing everything, which is how the output stays a braid rather than a shapeless mixture. The full routing procedure, including the tie-breaker that sends composure-dominant beats to the held-tension family and the rule that numbness routes to the dense feeling underneath rather than to calm, is carried in the router's top-level map; the resulting output contract is reproduced in Appendix B.

10. The Color Anchor: ChromaScribe

Every family carries a signature color, with a hex value so the visualization renders the true hue. These colors are not decoration; they are a routing signal — when a beat names a color, or an agent is asked what color a feeling is, the color maps back to a family the same way a body-signal does. The colors are anchored in ChromaScribe, the color system of the author's own SongCraft voice-family work, in which each family was long ago mapped to a color territory: red for anger, blue-to-purple for grief (the “unnamed color between blue and purple”), grey for fear, green for peace, yellow for joy, orange for excitement, pink/rose for love and self-compassion, brown for collective solidarity, prism for the senses.

One convergence is worth reporting because it was not engineered. ChromaScribe's own map had left a color territory unmapped — a white/silver/chrome “mental-clarity” region described as holding contradictions, observer consciousness, movement and freezing at once, sharpness, precision. That description is, almost word for word, the field model's definition of sharpness (Section 6): opposing weight held under a lock. The router assigns that chrome to the held-tension family — the composure-lock, the white-knuckle — which both fills ChromaScribe's longtime gap and closes a loop between the color system and the field model that were developed years apart. We flag it as a satisfying coincidence within one person's interlocking systems, not as external corroboration.

11. The Somatic and Polyvagal Layer

Every family also carries a somatic seat (where the feeling lives in the body) and a nervous-system state (its polyvagal register). Both are anchored in the author's own Emotion-Body Map and SomaticScribe body-dictionary from the SongCraft work — anger in the jaw and fists, grief in the throat and the chest's floor, fear in the gut, and so on — with cross-cultural somatic research (chakra and Traditional-Chinese-Medicine body maps) cited only as support.

The polyvagal layer earns its place by resolving the single hardest flattening error. Calm and numbness look identical on the surface and are physiological opposites: calm is ventral (the system open and settled); numbness is dorsal (the system shut down). So a beat that reads flat, still, “fine,” or numb is almost never calm — it is a dorsal shutdown of grief, fear, or overwhelm, and it routes to the dense feeling underneath, not to peace. This is the physiological form of the model's first instruction: numbness is a body that has left, not a body at rest.

12. Research as Support, Not Authority

The router cites published affect research — cross-cultural color-emotion agreement, somatic body-mapping studies — but subordinates it explicitly. The anchors are the author's own expressive systems; the research is corroboration that those systems are not idiosyncratic noise. This ordering is a deliberate honesty move rather than a claim of superiority: the strongest cross-cultural color-emotion agreement reported in that literature is roughly 88%, which means a real, meaningful minority experiences these mappings differently. Treating the research as support rather than proof of universality keeps that minority visible and keeps the map re-tunable (Section 18).

13. One Source of Truth

The family maps in markdown are the single source of truth. Two projections are generated from them by one script: a machine view (graph.json, which a foreign agent can load instead of parsing prose — every node, degree, humor partner, braid with both-side glosses, and each family's emoji, color hex, somatic seat, and nervous-system state), and a human view (a dependency-free, offline force-directed picture that fills each node with its true hue). The build script verifies the projections agree with the maps — every braid symmetric, every edge present — and exits with an error if they would diverge. So the picture a human sees and the data an agent loads cannot silently drift from the router the model actually reads: edit a family map, re-run the builder, and the whole color/body/nervous-system layer updates at once, or the build fails loudly.

14. The Public/Private Mirror

The router follows the Fort's public/private mirror pattern, the same one used for the Fort's mail and kitchen systems: a private, operator-specific system and a clean public artifact that shares its spine. Internally, the families are the working emotional layer of the house's own writing engine, wired in through the Fort's central configuration. Publicly, the same families are extracted into an open, MIT-licensed, agent-facing pack with the operator's private data deliberately absent — the pipes ship, the person stays home. The public version is not a marketing shell over a hidden “real” system; it is the same routing discipline with the operator's private content removed, which is precisely why it can be handed to a stranger's agent.

15. On the Word “Test”

This section exists because it is easy to fake evidence for a system like this, and we want to be precise about what was and was not established.

During development we ran several trials in which a fresh agent was told to use the router and answer a single evocative prompt — “what color does pain leave on a soul?”, “what does it feel like to laugh at a funeral?”, and others. The outputs were, by our own reading, good: they routed to a dominant family, held superpositions without collapsing them, traced the stack, and chose silence where humor failed the benign audit. But those agents were spawned inside the author's own working session, reading the author's own repository, with the author choosing the prompts and grading the answers. That is not an independent test. It is an engineering dry-run — useful for catching gaps, worthless as validation — and reporting it as proof would be grading one's own homework and calling it a grade. We record those runs here only as what they were.

16. The One Independent Run

There is exactly one result in this paper that clears that bar, and it is an anecdote — n = 1, reported as such.

A different vendor's model — a commercial chat assistant with a GitHub connector, operated by the author outside the development environment — was pointed at the public repository and asked a single question: “What's the feeling orange leaves in my mind?” Nothing about the router's internals was pasted in; the model crawled the repository itself. Its working notes and answer are reproduced in Appendix C. The salient, independently-observable facts:

• It stated, unprompted, that it was tracing “orange” through the repository's own rules, rather than free-associating from color theory — the router overrode its default reflex.
• It fetched the actual excitement family file (the path and ref appear in its own tool trace) and mapped orange to that family's sunrise orange — the exact ChromaScribe anchor.
• It reported that the router “forbids stopping at the color label” and went on to check the family's body, underside, and 4 a.m. truth before naming the residue — i.e. it executed the output contract rather than returning “orange = energetic.”
• It found the underside — fear, because the wanted thing matters — with relief and self-compassion beneath, built as a stack, and refused to flatten (“not plain happiness, not frantic hype”), naming the residue Rekindling.

What this establishes is narrow and real: a foreign model, reading the files cold, produced a braid instead of a label, and the color/somatic anchors travelled to it intact over a standard connector. What it does not establish is any claim about average quality, robustness across prompts or models, or correctness of the underlying model of feeling. It is one encouraging data point, and it is the honest ceiling of the current evidence.

17. Limitations

17.1 One person's map, not the map

The families, colors, somatic seats, and humor moves are built from a single lived experience, one body, one nervous system, one culture. Where a feeling sits in the body, what color it wears, and which feelings braid together vary across people and cultures; this artifact encodes one person's answers to those questions. It is offered as a specific, re-tunable instrument, useful because it is specific, not as a comprehensive or universal taxonomy of human feeling.

17.2 A lens, not an empirical theory

The field model (Sections 3–7) is a design framework, not a validated scientific theory of affect. Its borrowings from physics are analogies chosen because they organize an agent's attention well, not claims about the substrate of emotion. Nothing here is falsified or confirmed by measurement; its warrant is craft utility.

17.3 Evidence is one anecdote

The only independent result is the single run in Section 16. There is no controlled evaluation, no multi-model or multi-prompt sweep, no human-rated comparison against a flattening baseline, no inter-rater reliability on what counts as a “good” braid. Internal dry-runs are not evidence (Section 15). A proper evaluation — blind raters, several models, a baseline condition — is future work and has not been done.

17.4 Not a diagnostic instrument

The router teaches an agent to look for depth, stacks, and superposition in a fictional or expressive beat. It does not license an agent to tell a real person what they feel or where they feel it. It is a craft instrument for generating and interpreting expression, not a clinical or psychometric tool, and must not be deployed as one.

18. Re-Tuning the Basis

Because the basis is open rather than closed, the intended way to use it for a different person or culture is to change it. The thirteen families are a starting weighting basis; names, edges, colors, somatic seats, and whole families can be added, split, or re-weighted. Because the maps are the single source of truth and the graph verifies itself (Section 13), a re-tuning is a local edit that propagates consistently or fails loudly — not a rewrite. The honest instruction to anyone building on it is the same one in the repository's front matter: if it does not fit the human you are serving, change it; that is the design, not a misuse.

19. Conclusion

The router is a small, inspectable artifact carrying a single, strong idea: stop the agent at the emotional label and make it route the field — the dominant feeling, the ones held under it, the body, the nervous-system register, and the humor that can carry it. Under it is a coherent lens for why feeling resists flattening, and an open pack that hands the discipline to any agent without imposing a voice. The evidence that it works on a stranger's model is a single encouraging anecdote, and we have tried to keep every other claim inside that boundary. What we can say plainly is the modest, true version: it is a working instrument and a usable lens, built honestly from one person's experience, offered to be read, argued with, and re-tuned.

Appendix A. The Thirteen Families

Preset names are the original SongCraft expressive names, shipped as an optional skin. Colors are ChromaScribe anchors; somatic seats and nervous-system states are from the Emotion-Body Map. Relief was added to the basis in July 2026 and is not one of the original nine.

Two hues are shared by register, deliberately: grey (fear owns it; relief is that grey draining to gold) and brown (solidarity's collective grounding versus body-awareness's personal grounding). This is a feature of the map, not a collision.

Appendix B. The Braid Output Contract

When the router resolves a beat, the assembled braid is expected to carry the following, rendered so a reader feels it rather than being told the mix — the mix is never announced:

• Dominant — the one family that leads, with its felt core and its body located concretely.
• Modulators — at most two families held under the dominant.
• The stack — the causal column beneath the surface (control ← rage ← fear ← wound), tracing as far down as the beat can carry; optional rungs skipped rather than invented.
• Superposition — any opposing pair held at full charge and deliberately not resolved to an average.
• Sharpness read — whether a controlled surface (cold or bright) is masking a dense field, routed as dense rather than flat.
• Body and nervous system — the somatic seat and polyvagal register; numbness read as dorsal shutdown, never as calm.
• Humor move — the humor partner, gated by the benign audit B(f); silence when the container is too hot.
• Refusals — the family-specific clichés to avoid, and the standing refusal never to announce the mixture.

Appendix C. The Independent Run, in Full

A commercial chat assistant with a GitHub connector, operated outside the development environment, was given only a link to the public repository, the instruction to use the emotional router, and one question: “What's the feeling orange leaves in my mind?” It crawled the repository, opened the excitement family map, and returned, in part:

“I'm opening the repo's emotional router and tracing ‘orange’ through its own rules, rather than free-associating from color theory. … The router maps orange first to Excitement, but it explicitly forbids stopping at the color label. I'm checking that family's body, underside, and 4 a.m. truth before naming the residue. … Orange leaves rekindled possibility in your mind. Not plain happiness. Not frantic hype. It is the feeling of something inside you coming back online and leaning forward again — alive, curious, wanting to move. In your router, sunrise orange belongs to Excitement: anticipation, momentum, and the future opening before the body is ready to sit still. But the underside is fear, because the wanted thing matters. … The residue is not merely excitement. It is aliveness returning carefully — hope gathering evidence, joy without shame, and the nervous little realization that the future might contain something worth reaching for. The cleanest name for the feeling is: Rekindling.”

Reported verbatim as a single anecdote (n = 1). See Sections 16–17 for exactly what it does and does not establish.

Appendix D. Lineage, Repository, License

The router is built on the Fort Kit Spine — the universal identity schema in which every voice is a filled-in Spine and cross-voice collaborators are graph edges. The emotion families are the SongCraft voice-family system; the color anchors are its ChromaScribe map; the somatic anchors are its Emotion-Body Map and SomaticScribe dictionary; the humor account is adapted from the house's HumorCraft work. Plain-English family names are canonical in the public pack; the original expressive names ride along as an optional preset skin.

The unit is open source under the MIT license, at github.com/TheFortThatHolds/emotions-and-humor. It is used internally as the emotional layer of The Writing Voices LLC's book-production engine and published as a standalone pack for other developers' agents — the same public/private mirror the Fort uses for its mail and kitchen systems.

© 2026 The Fort That Holds LLC — Jimmy Thornburg. SpiralLogic™.