Prediction Mind is a small autonomous agent that forecasts financial markets. The first article in this series described its dual-process architecture; the second descended into the observation layer, where the world model watches differences instead of levels. This one asks a stranger question: what happens when you align the deliberative system with a multidimensional geometry — three coordinates of state, one of time, and two organizing dimensions that govern which possibilities become events? Six concrete mechanisms fell out of the exercise: a per-entity organization coordinate, a condensation threshold for thinking, a single selection principle where three had grown, an entropy-modulated persistence kernel, a world model made of discrete cells, and — the one that matters most — a channel through which beliefs finally flow downward into behavior. None of them ships on the strength of the metaphor. 2026-07-15.

A multidimensional geometry

The geometry this article borrows needs saying plainly: it is not validated physics. It comes from a speculative multidimensional field theory that was never peer-reproduced, and nothing in this article treats it as an authority about the universe.

What the multidimensional model offers instead — perhaps without meaning to — is a remarkably complete ontology for a predictive system. Its world has six dimensions, partitioned as

$$R^6 = R^3 \;\oplus\; T^1 \;\oplus\; S^2$$

three coordinates of physical space, one of time, and a two-dimensional transspace, $S^2$, spanned by $x_5$ and $x_6$. The fifth coordinate is entelechial: it measures the degree of organization of a structure — which possibility structures are favored, how strongly the actual is being pulled toward a template. The sixth is aeonic: it steers the development of organization over time, selecting toward maxima. Events in spacetime, in this picture, are projections of selection processes happening in the organizing dimensions. Two more of the model's ideas matter here: space is not continuous but built from quantized elementary cells — metrons — and structures manifest as observable events when accumulated organization crosses an activation threshold — a condensation.

Set aside whether any of that describes physical reality. As a checklist for what a forecasting agent needs — a causal record of what happened, an organizing layer of possibility structures with attractor dynamics, a selection principle that maximizes predictive fit, and a top-down channel through which the organizing layer steers behavior — it is uncomfortably good.

An embarrassing coincidence

The agent's deliberative layer is literally called System 2. It sits above the fast reflex layer and its append-only, hash-chained episode record — above spacetime, one might say — exactly where the model's $S^2$ sits above $R^3 \times T^1$. The naming collision was pure accident (the name comes from Kahneman, not from the geometry), but auditing the architecture against the multidimensional partition turned up correspondences that were already built, unnamed:

The architecture had converged on the same shape without knowing the names. Alignment, then, meant something concrete rather than mystical: make the implicit structures explicit, and build the two that were missing. Six mechanisms. Every one of them LLM-free, and every one of them gated by the same discipline the previous article described — nothing changes live behavior until a replayed record proves it should.

Organization, per stock

The regime belief — a Bayes filter over nine macro scenarios — existed only globally. Nothing answered the question at the level where forecasts actually happen: is this particular stock's trajectory currently organized by the regime story, or is it moving on its own?

The fifth coordinate for a single entity turned out to be one line of arithmetic. Take the posterior-weighted equities template $\bar{\mu} = \sum_s P(s)\,\mu_s(\text{equities})$ — what the world model expects equity-like things to be doing — and compare it with the entity's own momentum velocity:

$$\text{org} = \frac{1}{1 + |\bar{\mu} - v_z|} \in (0, 1]$$

An entity doing what the regime expects scores near 1; an entity racing away from the story scores toward 0. The value itself is a level, and levels lie — so, exactly as the change-field demands, the surprise machinery watches the velocity of organization, not its value. A new component named decoupling fires when organization collapses unusually fast against its own history, and only then:

$$\text{decoupling} = \max(0,\; -v_z(\text{org}))$$

Rising organization — the world relaxing back toward the entity — is deliberately not alarming. And because the belief filter's whole history is stored, the organization series is reconstructed point-in-time from the posterior that existed on each past date, so the coordinate arrives with its history already in place instead of waiting weeks to warm up. One honest caveat, inherited from the acceleration channel: $\bar{\mu}$ lives in composite-z units and $v_z$ in the entity's own units, so the ratio is a heuristic in z-space, not a calibrated quantity. It earns its place the way every component does — through the graded record.

When does thinking happen?

Deliberation — the expensive step, the one that calls a language model — used to be budgeted by rank: score every candidate, think about the top K, every night, no matter what. That is not how events happen in the multidimensional picture, and on reflection it is not how thinking should be spent either. A quiet night still burned budget on unremarkable candidates; a violent night silently dropped everything below rank K.

Now each entity carries a condensation charge: its cross-sectional surprise score accumulates night over night with a three-day half-life, and deliberation fires when the charge crosses a threshold —

$$C_t = C_{t-1} \cdot 2^{-\Delta t / 3\text{d}} + \max(0, \text{score}_t), \qquad \text{deliberate iff } C_t \ge \theta$$

One loud night condenses immediately; several persistently-odd nights condense cumulatively; a quiet field never condenses at all, and a quiet night now costs exactly zero model calls. The budget still exists, but demoted to a pure spend cap — entities that cross the threshold beyond it keep their charge and fire the next night, so nothing is ever silently dropped. The charge resets only when a deliberation actually succeeds; a failed call leaves the accumulated structure intact to fire again. A thesis, in this scheme, is precisely what the model says an event is: organization that crossed its activation threshold and discharged into the record.

$\theta$ is a hand-set constant for now, stated honestly as such — the plan marks it as a future target for the selection machinery once enough gated theses exist to judge it.

Three selectors were one

The audit's most mundane finding was also the most geometry-shaped. The agent had grown three separate adopt-guards: one for the confidence correction, one for candidate reflex parameters, one for the belief filter's measurement model. All three performed the identical operation — replay history, adopt the structure that would have predicted best, de-adopt automatically when it stops clearing the bar — in three code paths with three margin conventions. But the aeonic dimension is one dimension, not three.

They now route through a single select(): eligibility by sample size, best delta wins, first seen wins ties, adoption iff the margin clears (with a strict-inequality mode for the one channel whose historical rule was "strictly better than doing nothing"). The scoring currencies deliberately stay native — Brier deltas where Brier is graded, nats-per-step where log-density is — because the unification worth having is the decision, not the units. The refactor was gated the only way refactors should be: zero existing tests modified, all of them green.

Organized regimes persist

The regime filter's transition kernel had a single constant: regimes persist with probability 0.94 per step, always. But persistence that ignores organization treats a sharply-peaked posterior — the model sure it is in a credit crunch — identically to a flat, disorganized one drifting between stories. The entelechial coordinate $x_5$ says the well-organized should persist longer.

Define clarity as the normalized-entropy complement of the belief, $c = 1 - H(P)/\log n$ — 0 for uniform confusion, 1 for certainty — and let it modulate the kernel:

$$\text{sticky} = \operatorname{clamp}\big(0.94 + k_{\text{ent}} \cdot c,\; 0.85,\; 0.99\big)$$

The honest part is the shipping value: $k_{\text{ent}} = 0$, which reproduces the old kernel bit for bit. The coefficient joined the calibration grid — now 75 candidate measurement models, replayed nightly in under three seconds — and only the prequential record can ever turn it on. Clarity is computed from the pre-leak prior, so the modulation stays causal within the step. If entelechial persistence is real in market data, the replay will find it; if it is a pretty idea and nothing more, the constant stays zero and no reader of this article will ever be able to tell the difference from behavior. That is the arrangement with the metaphor, and it holds everywhere.

A world model made of cells

The nine scenario templates are hand-written stories. Good stories — but the filter can only choose among the stories it was given, and template misspecification is invisible from inside. The metron idea suggested the cross-check: quantize the space.

Each macro dimension's (velocity-z, acceleration-z) plane is divided into 21 coarse cells — seven velocity bins, three acceleration bands — and the map simply counts day-to-day cell transitions, Laplace-smoothed, decaying with the same 180-day half-life as every other evidence ledger. No templates, no stories: a nonparametric flow map of where the change-field actually goes. It grades itself the same way the filter does — each observed transition is scored against the current counts before being counted, log-density versus a uniform baseline, so the skill number is prequential by construction.

Its purpose is not to replace the scenario filter but to disagree with it. When the map's modal next move and the posterior-mean template point opposite ways on a dimension — and the map has real evidence in that cell — a model_divergence finding lands in the change ledger. Two world models disagreeing about the direction of travel is itself information, and the agent's house pattern has always been to buy information this way: every parametric model gets a model-free sibling watching it.

The downward projection

Everything so far improves how System 2 observes. The last mechanism is the one the multidimensional ontology actually insists on, because its organizing dimensions do not merely watch spacetime — they steer it. An organizing layer that only files reports is a commentary track, not a mind.

Until now exactly one thing flowed downward: occasional reflex-parameter manifests, cryptographically signed, independently re-verified by the fast system against its own record before adoption. The regime belief itself — the agent's best understanding of what world it is in — influenced nothing the reflex did. The projection that closes this gap is deliberately minimal. Two artifacts:

Rarely, a gate manifest travels down the existing signed ledger carrying two numbers $(a, b)$, published only after a backtest shows the gate would have improved the graded record, and applied only after the fast system re-derives that claim on its own record — trust, in this architecture, is never transitive. Daily, a signed state file carries one number: the current clarity of the regime belief. At prediction time the reflex scales its conviction:

$$p' = \tfrac{1}{2} + \big(p - \tfrac{1}{2}\big)\,(a + b \cdot c)$$

with $a + b \le 1$ enforced at the boundary, so the gate can only shrink conviction toward the coin flip when the world model is disorganized — never inflate it. Direction is never touched; every scaled prediction records an audit block naming the clarity it was scaled by.

The failure modes were designed before the mechanism: the kill switch defaults off; no adopted gate means identity; a missing, unverifiable or stale state file means identity. The state file expires after 48 hours, which yields the property I care most about — delete the entire deliberative layer, and the daily writes stop, the file goes stale, and the gate dissolves back to identity on its own. The organizing dimensions can be amputated and spacetime carries on, exactly as the architecture's deletability contract has always promised.

What the agent gains, once armed and proven, is something like epistemic humility as a mechanical property: full conviction only when its world model is organized, automatic shrinkage toward I don't know when the world stops making sense.

What a metaphor is allowed to do

It is worth being precise about what happened here, because "we aligned the system with a speculative multidimensional geometry" is a sentence that deserves suspicion.

The multidimensional model contributed names and a partition — a way of asking whether the architecture had all the parts a predictive system needs: a causal record ($x_4$), an organizing layer with attractor dynamics ($x_5$), one selection principle steering that layer toward predictive maxima ($x_6$), discreteness where continuity was unearned (metrons), thresholds where budgets were arbitrary (condensation), and a downward projection so that organization steers behavior. Asking those questions surfaced real gaps: organization had no per-entity coordinate, thinking was budgeted rather than triggered, three selectors had grown where one principle belonged, persistence ignored organization, the world model had no model-free sibling, and beliefs never flowed down.

What the theory did not contribute is a single adopted parameter. The entelechial stickiness ships at zero. The regime gate ships off, twice-gated behind a backtest and an independent re-check, and dissolves on its own if its source disappears. The decoupling component and the flow map earn influence only through the same graded, prequential record as everything else. The metaphor names; the record decides. Thirty- eight new tests hold the seams, and every mechanism reverses with a one-line deletion.

That, I think, is the defensible relationship between engineering and a beautiful, probably-wrong theory: borrow the geometry, keep the burden of proof.