Domestication Space (psi) diagram

Figure 1. “Domestication Space (ψ)” — a 3D affinity field with an explicit target vector and a topology threshold. (Interactive)

Figure 2. Static rendering of Domestication Space (ψ) showing bounded cube (0–1.1), sample species plotted as 3D coordinates, Euclidean threshold edges (≤ 0.22), and the explicit aspirational target point at (1.1, 1.1, 1.1). This diagram isolates structural assumptions independent of interactivity.

The moment you define an axis, you have defined a value.

“Domestication Space (ψ)” looks technical: a cube, a distance metric, a threshold graph, a synthetic generator. But it is not primarily about animals. It is about what happens when you convert a story about selection into a coordinate system, and then let geometry decide what is good.

In the interface above, each species occupies a point in a 3D field defined by Bond, Utility, and Tolerance. There is an explicit ideal at (1.1, 1.1, 1.1). Distance to that point becomes an “affinity score.” Proximity below a fixed threshold becomes a topological edge.

Nothing about that is neutral.

The cube is a contract. It says: domestication can be decomposed into three stable dimensions; we can measure them; we can rank species by Euclidean distance; and we can treat the resulting scalar as meaningful. If that sounds abstract, consider how often the same move occurs in AI systems: define latent dimensions, compute distance, assign score, deploy.

ψ makes that move visible.

1. From narrative selection to objective function

Domestication is usually told as narrative. Wolves linger near human camps. The less aggressive individuals tolerate proximity. Over generations, those temperaments become amplified. Eventually: dogs.

The mid-20th-century fox experiments make the point directly. Selecting for tameness did not merely reduce aggression. Within a few generations, researchers observed morphological changes: floppy ears, altered coat coloration, modified skull shape, and shifts in reproductive timing.^[1]

This bundle of correlated changes is often discussed under the umbrella of “domestication syndrome,” and one leading hypothesis connects it to neural crest cell development and its wide-reaching developmental influence.^[2] The important design lesson is not the biological mechanism. It is the structural fact: optimize one dimension hard enough in a coupled system, and you will move others whether you intended to or not.

Selection is never one-dimensional in effect, even if it is one-dimensional in intention.

In ψ, we do not pretend that Bond, Utility, and Tolerance are biologically precise constructs. They are abstractions. But the abstraction itself mirrors the domestication move: isolate a dimension (or three), declare it central, and measure progress against it.

The fox experiments show what that implies. Once you formalize a selection criterion, you inherit every hidden dependency that criterion activates.

In AI systems, the parallel is direct. When you optimize a model against a single loss function, you rarely optimize only that surface-level quantity. You are reshaping internal representations, biasing outputs, redistributing errors. The difference is scale and opacity.

ψ is intentionally small and inspectable. It shows you the axis labels. It shows you the metric. It shows you the target. Most deployed systems do not.

2. The metric is the ethics

In the demo, the affinity score is derived from Euclidean distance to a fixed target. That decision embeds at least three normative assumptions:

Each axis contributes symmetrically.
Tradeoffs are linear and continuous.
The “ideal” is a single point, not a region.

If we changed the metric to Manhattan distance, the ranking could change. If we weighted Utility twice as heavily as Bond, the ordering would shift again. If we replaced a point target with a volume (e.g., “anything with Tolerance > 0.9 is acceptable”), the geometry of success would alter.

These are not mathematical curiosities. They are policy decisions disguised as implementation details.

In embedding-based systems, the same structure recurs. Entities are mapped into vector spaces. Similarity becomes distance. Distance becomes proxy for meaning.

Interpretation, however, is layered on top of that geometry. Proximity does not guarantee semantic coherence, and interpretability work has shown how fragile these inferences can be.^[3]

The cube in ψ is deliberately literal. You can see the bounds. You can see the edges forming when species fall within a threshold of 0.22. That threshold is arbitrary. It could be 0.15 or 0.35. Change it, and the topology of “affinity” changes.

Topology is argument.

If two species are connected by an edge, the interface implies similarity. If clusters emerge, it implies natural groupings. But these are artifacts of threshold choice and metric structure. In real-world algorithmic systems, those artifacts can influence resource allocation, recommendation patterns, enforcement decisions, and public perception.

When opaque scoring systems operate at scale, especially in domains like credit, policing, or employment, the stakes of metric choice are not theoretical. As critics of algorithmic decision-making have argued, hidden objectives and inaccessible models can amplify inequities while appearing mathematically objective.^[4]

ψ is a toy, but it enacts the same chain:

Field → Metric → Score → Structure → Action.

The only difference is that here, the chain is visible.

3. The explicit ideal: (1.1, 1.1, 1.1)

The target point at (1.1, 1.1, 1.1) is intentionally exaggerated beyond the unit cube. It signals aspiration beyond observed data. No species in the baseline dataset naturally occupies that coordinate. The generator exists to synthesize one.

This mirrors a recurring pattern in AI product design: define an aspirational objective not fully grounded in empirical distribution, then use generative tools to populate that objective space.

The moment the “Synthesize Ideal Species” button is pressed, the system stops merely evaluating reality and begins producing candidates tailored to the objective. The generator is constrained to output a1, a2, a3 values between 0.98 and 1.1. That is not discovery. That is guided construction.

Once generation is coupled to scoring, optimization becomes self-reinforcing.

In biological domestication, the breeder selects from existing variation. In AI-mediated synthesis, we can generate variation directly in the objective space. That compresses evolutionary timescales from generations to milliseconds. It also magnifies the importance of the objective definition.

If the objective is flawed, the system will not drift slowly toward error. It will accelerate.

4. A controlled perturbation: what happens if Utility ×2?

The fastest way to test whether a metric is ethical is to perturb it.

Suppose we reweight the objective so that Utility counts twice as much as Bond and Tolerance. In formal terms, instead of minimizing:

d = √[(1.1 − a1)² + (1.1 − a2)² + (1.1 − a3)²]

we minimize:

d′ = √[(1.1 − a1)² + 4(1.1 − a2)² + (1.1 − a3)²]

To make this concrete, consider two baseline species in the current dataset:

Dog: (0.95, 0.85, 0.90)
Horse: (0.85, 0.95, 0.70)

Under symmetric Euclidean weighting, Dog’s stronger Bond and Tolerance keep it closer to the (1.1, 1.1, 1.1) target overall. Horse benefits from higher Utility but is pulled back by lower Tolerance.

If Utility is doubled, the gap narrows sharply. The penalty for Horse’s slightly lower Bond becomes secondary to its stronger Utility alignment. Depending on exact normalization, Horse can overtake Dog in affinity ranking.

Nothing biological changed. The ranking flipped because the coefficient changed.

That is the entire point.

When teams debate “what matters most,” they are debating weights.

Now watch what happens conceptually:

Species high in Utility but moderate in Bond rise in rank.
Highly bonded but less instrumentally useful animals fall.
Clusters near the Utility axis compress; those far from it fragment.

Nothing biological changed. Only the declared importance of one axis shifted.

This is precisely what makes metric design non-trivial. Rankings emerge from the metric, not the species.

In the current demo, weights are fixed and symmetric. That is intentional. It makes the ethical choice legible: symmetry is a stance. It encodes the claim that affection, usefulness, and tolerance are equally important.

If we expose weight sliders, disagreement becomes parameterized instead of rhetorical.

The lesson generalizes: in any embedding-based production system, the difference between “fair” and “distorted” may be a coefficient.

5. Figure 2: static geometry, visible assumptions

For non-interactive contexts, the static diagram below captures the same field structure.

Figure 2 shown above as inline SVG for binary-free publishing workflow.

Figure 2 makes three structural assumptions explicit:

Boundedness — the field is constrained between 0 and 1.1 on each axis.
Centralized aspiration — there is a single ideal target point.
Local similarity — edges form below a fixed Euclidean threshold.

The static rendering removes animation, synthesis, and hover interactivity. What remains is the argument.

If you were auditing this system without code access, this figure would be your starting point. It tells you where success lives and how similarity is constructed.

Many real-world AI deployments do not provide even this.

6. From toy to product: making objectives inspectable

Most AI products already operate on a hidden ψ. A latent space. A similarity metric. A ranking layer. A generation or selection mechanism. A threshold for action. What distinguishes responsible systems from brittle ones is not sophistication, but inspectability.

A real-world analogy clarifies the stakes.

In semantic search systems, documents and queries are embedded into vector space. Ranking depends on distance or cosine similarity. If you adjust the loss function during training to emphasize engagement signals over semantic coherence, the resulting space will cluster differently. Highly clickable but semantically weak items may drift closer to the “ideal” query region.

From the outside, the system still returns “relevant” results. Internally, the geometry has changed.

ψ makes that drift visible. It shows that a small weight change can reorder outcomes and rewire local neighborhoods. In production search or recommendation systems, the same shift can change what users see, what content spreads, and what creators are rewarded.

Vector spaces do not merely describe preference. They operationalize it.

Concretely, this stack includes:

A latent space.
A similarity metric.
A ranking function.
A generative or selection layer.
A threshold for action.

What is usually missing is inspection.

If we were to ship ψ as a production-grade component, three additions would be non-negotiable:

Objective logging. Every ranking decision logs the weight vector and metric configuration used.
Counterfactual replay. Users can recompute outcomes under alternate weight settings.
Topology diffing. Graph structure changes are tracked as thresholds shift.

These are not performance features. They are epistemic safeguards.

The structure of the metric determines the structure of the outcome.

Critiques of opaque algorithmic systems consistently identify the same failure mode: decisions appear mathematically justified, but the underlying objective is inaccessible or contestable only by experts.

ψ demonstrates the inverse posture. It says: here are the axes. Here is the metric. Here is the threshold. Change them and see what happens.

This is not just transparency. It is parameterized accountability.

7. Synthesis as acceleration

The “Synthesize Ideal Species” button closes the loop.

In traditional domestication, selection operates on existing variation. In ψ, generation occurs directly inside the objective field. The generator is constrained to produce coordinates near the ideal. That means we are no longer searching the space. We are filling it.

This matters because generative systems accelerate optimization. If the objective is misaligned, generation compounds the misalignment faster than selection alone would.

Embedding interpretability research reminds us that vector proximity does not guarantee semantic coherence. When generation and scoring share the same latent assumptions, feedback loops can form.

ψ does not hide this loop. It stages it.

The deeper claim is not about animals. It is about interface design under optimization pressure. If we accept that objectives are unavoidable, then the responsible move is to expose them, perturb them, and log them.

Field → Metric → Graph → Generation → Audit.

That is the minimum viable structure of an honest optimization system.

8. Alignment is metric design

The current alignment discourse often frames the problem as one of instruction-following, safety constraints, or reward modeling. But beneath those layers sits a simpler structure: a system optimizes something.

Loss functions in supervised training, reward models in reinforcement learning, preference datasets in RLHF — each encodes a weighted objective. The debate over alignment is, in operational terms, a debate over coefficients.

ψ is a deliberately small mirror of that structure. Bond, Utility, and Tolerance are not moral truths. They are declared dimensions. The system ranks according to their combination. When we doubled Utility, we changed what “good” meant without touching the data.

That is alignment in miniature.

In large systems, the dimensions are less legible and the coupling effects more complex. But the core move is identical: define what success looks like, optimize toward it, then observe second-order effects.

The fox experiments illustrate that optimizing one trait can produce unexpected morphological bundles. Modern AI systems reproduce the same structural dynamic at computational scale. Optimize engagement and discourse shifts. Optimize click-through and distribution patterns change. Optimize short-term retention and long-term trust may degrade.

The structure of the metric determines the structure of the outcome.

Alignment, then, is not primarily about finding the perfect rule. It is about exposing, testing, and iterating on the objective function itself.

Conclusion: ψ as minimum viable honesty

The Domestication Space demo does not solve alignment. It does not solve domestication. It does not even claim to measure reality with precision.

What it does is simpler.

It refuses to hide the objective.

It names the axes. It fixes the metric. It draws the threshold. It exposes the ideal. It allows perturbation. It couples generation to scoring in plain view.

Most deployed optimization systems already follow the same chain. They simply conceal it behind abstraction layers and scale.

ψ proposes a different norm: if a system ranks, it must reveal how. If it optimizes, it must expose the objective. If it generates, it must log the configuration that made that generation desirable.

Field → Metric → Graph → Generation → Audit.

This is not a slogan. It is a minimal contract between optimization and accountability.

Rankings emerge from the metric, not the species.

If we accept that, then the primary design question shifts. Not “What is the perfect output?” but “What is the declared objective, and can we test it?”

That is where serious alignment begins.

Contextual Recommendation

If you want to treat vector objectives as navigable structures rather than slogans, take the next step into a general-purpose inspection tool. Primary Design Co.’s JS Visualizer is the most direct continuation of this pattern: it assumes that structure should be traversed, not merely asserted.

Explore the JS Visualizer: https://primarydesignco.com/js-visualizer/

Footnotes

^[1]: Overview of the Russian farm fox experiment and behavioral selection effects, including American Scientist’s account of early canid domestication and fox selection results; and Lord, K. A., Larson, G., Coppinger, R. P., & Karlsson, E. K. “The History of Farm Foxes Undermines the Animal Domestication Syndrome.” Trends in Ecology & Evolution 35(2), 125–136 (2020).

^[2]: Wilkins, A. S., Wrangham, R. W., & Fitch, W. T. “The ‘Domestication Syndrome’ in Mammals: A Unified Explanation Based on Neural Crest Cell Behavior and Genetics.” Genetics 197(3), 795–808 (2014).

^[3]: Tennenholtz, G., et al. “Demystifying Embedding Spaces Using Large Language Models.” OpenReview (2024).

^[4]: O’Neil, C. Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy. Crown (2016), and related reporting on opaque algorithmic harm.

References

Lord, K. A., Larson, G., Coppinger, R. P., & Karlsson, E. K. (2020). The history of farm foxes and the role of behavioral selection in domestication. Trends in Ecology & Evolution, 35(9), 814–826.
Wilkins, A. S., Wrangham, R. W., & Fitch, W. T. (2014). The “Domestication Syndrome” in mammals: A unified explanation based on neural crest cell behavior and genetics. Genetics, 197(3), 795–808.
Tennenholtz, G., et al. (2023). Demystifying embedding spaces using large language models. OpenReview.
O’Neil, C. (2016). Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy. Crown.