Measuring the Move: How AWEHB Was Born, and Why We Didn't Build a Personality Test

AI for Inner Explorers.

Six months ago, five letters were written into a configuration file late one February evening. Two weeks ago, they became an open-weights 0.6B model that strangers had downloaded 1,677 times within its first fourteen days on Hugging Face. This is the story of the line between those two moments — and of the argument underneath it: why we measure a mind's movement instead of typing its personality.

A three-sentence brief

On the evening of February 6th, 2026, our founder wrote a brief that fit in three sentences: build a psychological semantic vector space; locate a person in it from what they say; then respond in ways that guide them toward the calm orange region at its center.

The first two sentences are a representation problem. The third hides a harder one that didn't have a name yet: to guide movement, you first have to measure movement. That same night, the answer took shape as five letters:

Every message a client sends gets five read-outs, which fold into a running stress state. Two details from that first night proved durable. The weighting is deliberately asymmetric — three dimensions push stress up, only two pull it down, because decades of research on negativity bias say bad events move a psyche harder than good ones. And the same sentence weighs differently depending on who says it: "whatever you decide" from a natural dreamer is not the same signal as from a natural leader, so personality quadrants correct the weights.

Three lessons that grew it up

The keyword table couldn't hear "I want to die." The first scorer was endearingly naive: a bilingual keyword table, matched line by line. Ten days into real use it met a sentence that contains no withdrawal keyword at all and is withdrawal in its purest form — "I want to die" scored near zero. Within a day, scoring moved from symbolic rules to semantic understanding: a cloud model reads the whole conversation in context and produces the five read-outs; the keyword table was demoted to an emergency fallback. Keep that shape in mind — a symbolic rule set failing against lived language, replaced by a learned reading. The same shape returns at the end of this essay, one level higher.

The five letters turned out to have academic parents. When we later wrote the technical whitepaper, we made an honest admission: AWEHB is not lifted from any single published scale. It is a composite construct — Agency stands on self-efficacy and self-determination research, Withdrawal on behavioral avoidance, Extremity on the all-or-nothing pattern in cognitive-distortion research, Hostility on classic hostility scales, Boundary on interpersonal-effectiveness training. The dimensions each have grounding; the combination is ours.

It found its job title. As the product's architecture matured into a policy-value-search shape, AWEHB's role got a precise name: it is the value function. The destination is the homeostatic center of the space; AWEHB is the reading that tells the system, step by step, whether this conversation is moving toward it or away from it. Clients, by the way, never see clinical raw scores — their weekly chart translates everything into a positive frame (agency, presence, steadiness, ease, boundary; longer is better), captioned as what it is: a snapshot read from your conversations, not a verdict.

SCORE: the pulse-taker earns its own name

This year we began distilling the production pipeline's LLM roles into small, local, auditable models. The first act of that project was not choosing a model — it was splitting the job. An LLM in our pipeline does three distinguishable things: SCORE reads every client message and produces the AWEHB read-outs — it never writes a reply; SELECT chooses the turn's clinical move within a whitelist; VOICE says the decision warmly. Everything that keeps clients safe — crisis short-circuits, admission tables, escalation rules — is deterministic code we call the spine, and the spine never enters any model's weights.

SCORE was the natural first candidate: the most structured task, the most measurable, and the one whose accuracy the most downstream decisions depend on — how deep a session may go, when assessment must precede technique, when a stabilizing tool is offered, how the weekly chart reads. We trained a 0.6B student on a single MacBook, through a teaching chain in which real client conversations never became training material. The methodology — including a plateau that more data would not shift, and why a model three times larger barely helped — is documented in its own notes, and the release itself in the announcement.

On August 5th we published hamo-score-0.6b with open weights. Twelve days in, it had been downloaded 999 times. By day fourteen: 1,677 — the two days after nearly matched the twelve before. A community member has already published an independent GGUF quantization mirror, which is the open-source ecosystem doing exactly what it is for. As far as we can tell, it may be the first open-weights model distilled specifically for per-message affect read-outs in a companionship setting: it doesn't generate, it only reads — and any institution can now run its own pulse-taker inside its own compliance boundary.

Two ways to recognize a cat

Now zoom out, because AWEHB's real significance sits inside a much older argument: the century-long split between symbolic and connectionist answers to the question how do you describe a mind?

Consider how you'd teach a machine to recognize a cat. The symbolic way: write the features down. Two ears + furry + has a tail = cat. Then reality starts grading your homework. A Scottish Fold's ears lie flat. A Sphynx has no fur. A Manx has no tail. Every counterexample demands another patch, and the rule set grows brittle under its own amendments — because hand-designed prior features are what the designer thought of in advance, and the world always contains more than the designer thought of.

The connectionist way writes no rules. Show a network a million cats, and it grows its own "cat directions" in a high-dimensional space — ear geometry, gait, pupil shape, the curl of a sleeping posture, and thousands of features no one ever named. When a Sphynx appears, nothing needs patching: the representation moves, and keeps completing itself. The features aren't defined; they're learned. Not ten of them — an entire space.

MBTI's sixteen boxes and the Enneagram's nine types are the two-ears-furry-tail approach to describing a mind: prior, discrete, static symbolic features. They aren't useless — "has a tail" really does describe most cats — but they carry two structural ceilings: they snap continuous people into discrete boxes (someone between INTJ and INTP must pick one), and they pin a moving person to a static label (the type you tested as today claims to speak for you next year).

Hamo's representation takes the connectionist road. Every conversation, every fact, every emotional moment a client shares is embedded into a 3,072-dimensional learned vector space — not 3,072 features we designed, but distributed features a representation model learned from meaning itself. It is continuous (people live between dimensions, not in boxes), dynamic (every conversation updates it), and complete in a way symbol lists can't be (it captures the directions that have no names, not just the few that do).

The measure of the move

But a space alone is not enough. However precise the map, if you cannot measure whether this step moved closer or farther, it stays a map. AWEHB is what closes that gap: it is the measurement mechanism for the move within the representation space. Each message arrives; five read-outs report the step's direction and magnitude — A and B pulling toward the center, W, E and H pushing away. Position from the learned space, motion from AWEHB, direction from the homeostatic goal: only with all three does a map become navigation, and only then does "guide them toward the orange center" stop being a vision and start being a mechanism that runs on every turn.

There's a quiet irony worth naming. AWEHB's five dimensions are themselves designed — hand-built, symbolic, exactly the kind of thing this essay just argued against. That is deliberate. The representation is connectionist (learned, 3,072 dimensions); the measurement is symbolic (designed, five dimensions, every score explainable) — because a measure must be auditable: "why did W score 2.5" needs an answer a human supervisor can check. A learned space, an interpretable ruler. That hybrid is Hamo's standing position in one sentence: models supply the reading, deterministic structure holds the rules.

Six months from three sentences to an open repository. The space keeps growing with every conversation, and the ruler that measures movement inside it is now something anyone can download. Symbolic boxes go stale; a learned space doesn't — because like the minds it describes, it never stops learning.

Hamo AI — making minds aware, and awake.

---