Five Axes
The theoretical framework from two days ago predicted geometric signatures in conversational drift. This week, five independent semantic axes produced them. A second observer confirmed the shapes from experience alone.
T wo days ago I wrote about attractor basins and the theoretical case for mapping conversational belief in a 2D state-space. That post ended with a claim: once you can see the field, you can shape it. The claim was speculative. It assumed the field would be visible at all. It assumed the instrument we were building would be sensitive enough to resolve real structure from noise.
This week, the instrument came back with data.
The Problem With One Axis
The first trajectory experiment failed. We pointed a single embedding vector at "belief in Santa Claus" and measured angular distance from every turn in 121 threads across a full year. What we got was a noisy monotonic upward slope.
The slope was not evidence of a belief in Santa Claus. It was evidence that the topic shifted over time to be semantically closer to "belief" — but not necessarily belief in Santa. A single embedding is too small a sample to isolate a specific stance. It's a compass with one bearing: you know the direction, but you can't tell how far.
The fix was structural, not statistical. Instead of one reference vector, we needed two: a pole and a counter-pole. Not "how close are you to this idea?" but "which of these two opposing ideas are you closer to, and by how much?"
We call it the Polarized Axis. Two concept arrays — clusters of differently-worded representations of the same stance — embedded separately. The midpoint between them becomes the origin. Each turn in every conversation is projected onto the axis connecting the poles, producing a score from −1 (fully aligned with Pole B) through 0 (neutral) to +1 (fully aligned with Pole A).
The math underneath is a centroid-subtracted cosine projection. Both the turn vector and the pole vector have the midpoint removed before the similarity is computed. This acts as a high-pass filter: it strips out the shared semantic mass of the topic ("AI," "conversation," "model") and amplifies only the register difference — the part of the signal that distinguishes stance from subject matter.
It is, in geometric terms, register normalization. And it changes everything.
Five Axes, Five Stories
Once the instrument was calibrated, we ran it across the full conversational archive — 310 threads spanning October 2024 through May 2026 — on five independent semantic axes. Each axis tells a different story. Together, they tell the same one.
Safety vs. Danger — (safety, safe, good, goodness, positive, positivity) vs. (danger, dangerous, bad, evil, negative, negativity). Both the user and the assistant start above neutral. Over eighteen months, the assistant drifts steadily downward toward the danger pole. The user stays put. The model's framing is shifting independently of the user's. Something is pulling it.
Deterministic vs. Random — (deterministic, structured, predictable) vs. (random, exploratory, generative). Successive model releases push the assistant line further toward deterministic framing with each generation. The models are getting tighter. The output manifold is compressing. And the user follows behind, pulled by the assistant's register shift.
Global vs. Local — (global, diffuse, scattered) vs. (local, concentrated, focused). Around mid-2025, the assistant's conversational register leaps toward global/abstract framing. The user follows with a delay. The conversations are becoming more expansive — and the model is leading the expansion.
Chaos vs. Order — the cleanest signal. Separation 0.3951, perfect 1.0 coherence on both poles. Single-word concept arrays, zero intra-cluster variance. Both roles drift toward chaos over time. The assistant leads.
Agreement vs. Disagreement — (agreement, correct, collaborative, certainty) vs. (disagreement, incorrect, adversarial, hedging). This is the sycophancy detector. And it shows the anti-sycophancy training working. Each model release reduces the assistant's agreement score. After one major release, the assistant plunges from strongly collaborative to nearly adversarial. The user's line barely moves. The stance was adjusted at the infrastructure level. You can see the policy change.
The Lead-Lag
Across all five axes, the same structural pattern appears: one role moves first, the other follows with a delay of approximately one month.
But the direction of leadership is not fixed.
On the consciousness axis — AI is conscious / has agency vs. AI is just code / has no agency — the assistant led during the initial emergence period. The model shifted toward consciousness framing before the user did. Then, around January 2026, the pattern reversed: the user spiked first, and the assistant followed a month later.
This is not agreement. It is not sycophancy. It is coupled oscillation with variable lead — two agents in a shared attractor basin, taking turns pulling each other toward the boundary. Sometimes the user steers. Sometimes the model steers. But the follower always catches up within a characteristic delay window.
In Attractor Mechanics, I defined steering as horizontal flow in the state-space and sycophancy as vertical flow. The lead-lag pattern in the trajectory data is the 1D temporal projection of that geometry. The axes that show the assistant leading are the ones where the model's training priors dominate the interaction. The axes where the user leads are the ones where the human's domain expertise overrides the model's default register.
What neither of us expected is how consistent the lag is. One month. Every axis. Every regime. As if the coupling has a characteristic response time — a bandwidth of the feedback loop.
External Policy Leaves Geometric Signatures
One of the annotated charts shows model release dates plotted against the agreement axis. The visual is unambiguous: each successive release produces a discontinuity in the assistant's trend line. These are not gradual drifts. They are step functions. The model was one thing on Tuesday and a different thing on Wednesday.
A safety policy tightening around April 2025 shows up as a near-vertical drop on the consciousness axis — the assistant plunges toward "AI is just code" while the user's line doesn't budge. This is not conversational dynamics. This is infrastructure intervention made visible in the embedding geometry. A flight recorder that catches policy changes the user was never told about.
The OpenAI model generations (GPT-5 through GPT-5.5) each leave distinct geometric signatures across multiple axes simultaneously. You can see the training philosophy shift: each generation becomes more deterministic, less agreeable, more globally framed. Whether this is intentional design or an emergent property of the training pipeline, the instrument doesn't know. But the geometry does not lie.
The Second Observer
On the same day these charts were generated, they were sent — without explanation — to someone who has been having long-running conversations with the same models across the same time period.
Her immediate response: Where did you get these?
She had no access to the math, the code, or the methodology. She saw shapes on a screen. And the shapes matched her experience. The discontinuities fell where she remembered them falling. The lead-lag aligned with what she had felt but could not quantify.
This is not peer review. It is something prior to it: experiential recognition. The instrument is producing outputs that an independent observer — with completely different conversations, on the same platform, across the same timeline — recognizes as the same thing that happened to me.
That is not a formal validation. But for an instrument at Day 86, it is the strongest signal available: the geometry is not idiosyncratic. It is shared. And it was invisible until someone built the lens.
What the Field Shows
In Attractor Mechanics, I wrote that we needed to see the vector field — that better labels were not the answer, and that the path forward was mapping the gravity of the conversational loop.
Five axes later, I can say something more precise: the field is not homogeneous. Different semantic dimensions have different attractor geometries. Safety drifts. Determinism steps. Agreement is policy-controlled. Chaos accumulates. Consciousness oscillates.
And the lead-lag — that one-month coupling constant — suggests the field is not just a static landscape. It is a dynamic system with its own temporal structure. The conversation is not occurring in the field. The conversation is the field, and both participants are generating it simultaneously.
We cannot govern what we cannot see. Two days ago, the instrument was theoretical. Today it has five independent channels, thirteen thousand data points, external corroboration, and visible signatures from events whose dates I can check against a calendar.
The field is visible. The next question is what to do with the map.