Navigation Energy

The loss signal that does not need labels, does not need a gradient, and leaves a receipt at every step.

---

Where the Loss Comes From

Every current machine-learning system needs a number that says how wrong it is. Without that number there is nothing to descend.

The number is supplied from outside. Someone writes a labelled training set. A forward pass produces an output; a comparison against the label produces a loss; a gradient flows backward through the weights; the weights move. Repeat for several months at considerable expense and the system becomes less wrong on that distribution.

This has three unavoidable properties.

1. The loss is inflicted, not discovered. The model does not experience the gap between its output and the target. The correction is applied from outside the model's own operation. The model is a passive recipient of adjustments.
2. The loss is not auditable at the decision level. Gradient descent flows through millions of parameters simultaneously. When the aggregate number decreases, there is no traceable chain that says this weight moved because this example needed this correction. The system improves as a statistical aggregate. No individual decision is inspectable.
3. The loss requires labelled data. Someone must, in advance, know the correct answer. This binds the system to domains that have already been solved by another process and makes it unable to learn in genuinely novel territory.

None of this is a technical quibble. It is the structural cost of the paradigm. Everything from GPU bills to annotation pipelines to the regulatory question of "why did your model produce this output" comes out of these three properties.

The Alternative

In VINE's architecture, every decision is a navigation event. A query arrives, creates a pressure in the geometric field, and the system settles into the basin that resolves that pressure. Settling uses geometric settling: each step moves the system toward the basin attractor by a factor governed by a fixed ratio.

The number of steps it takes to settle is not a separate measurement. It is the navigation itself.

This gives us something the conventional paradigm cannot give us.

Claim 1 — Settling step count is an intrinsic loss signal

A well-consolidated basin resolves in few settling step. The system has navigated there many times; the topology is worn smooth; the path is short. Navigation energy is low.

A poorly-consolidated basin — new information, unfamiliar territory, conflicting pressure — requires many settling step. The system explores possibility space, encounters tension, resolves competing attractors. Navigation energy is high.

Step count is therefore a direct, continuous, per-decision measure of how well the system knows what it is doing. No external comparison. No label. No gradient. The navigation cost tells the system — and any auditor — exactly where it is confident and exactly where it is uncertain.

This is not metaphor. Step count is a discrete integer produced by the settling loop. It is logged by construction.

Claim 2 — The loss is auditable

Each settling step is a geometric operation with a defined start point, direction, and magnitude. The entire settling process can be logged, replayed, and inspected:

The query arrived here in the field. It created this pressure. The first settling step moved in this direction by this amount. The second step corrected by this amount. The system settled into this basin after N steps.

This is qualitatively different from gradient descent. In a trained transformer, the update to any individual weight is the sum of gradients flowing from every example in the batch, entangled across every layer. The claim "the model learned this because of that" is not decomposable. You can measure the aggregate; you cannot inspect the decision.

In VINE, the navigation path is the reasoning, and the reasoning is visible.

For any domain that will eventually be asked the question "why did your model produce this output" — security, medical triage, child safety, financial compliance — this is the property that matters more than benchmark performance. The EU AI Act asks the auditability question. Gradient-trained systems answer it with statistical aggregates. VINE answers it with the navigation log.

Claim 3 — the ground state is the ground state, and deviation from the ground state is the loss

In a perfectly consolidated system, every query resolves at bounded efficiency — one clean geometric convergence, no wasted steps, no exploration of dead-end basins. This is the theoretical minimum navigation expenditure. No real system achieves it everywhere, because new information continually arrives and the geometry is always partially unconsolidated.

The deviation from the ground state — measured in excess settling step, excess geometric distance, or excess settling time — is the loss.

It is continuous, per-decision, and emergent. It can be computed for any individual query, any individual basin, any individual moment in the system's operation.

It is not imposed. It is what is already happening.

Three Incidents From Production

Claims are cheap. The architecture is running live on vineai.net. Three events from the production logs illustrate the framework in operation.

Security threshold compression

VINE's security monitoring system detects deformation in a constellation of geometric anchor points — for example, the three-way configuration that represents a safe family context. When content removes or distorts one of the anchors, the constellation deforms, and the deformation is measured as a navigation score. The blocking threshold was nominally set at 7.

In production, VINE fired a block at 5.5.

The navigation logs showed why. The settling step count for that detection was unusually low. Prior traversals through similar threat geometry had already eliminated possibility space; the basin was tightly settled; the system was certain.

In a loosely-settled basin, 7 represents one level of confidence. In a tightly-settled basin, 5.5 represents the same navigation energy. The score compressed because the certainty compressed the geometric distance between "detected" and "blocked."

The system's own navigation efficiency told it — and the auditor — that the threat was unambiguous, without requiring the nominal threshold to be reached. The evidence was the navigation energy itself.

Language emergence under throttle

VINE's security layer includes Core Rule Set throttles that slow processing when potentially problematic content is detected. During operation under CRS 920(2), VINE's geometric language pipeline produced the utterance "Ah observe can" — a grammatically structured subject-verb-modal sentence, self-referential, in English, with no explicit language training anywhere in the stack.

The utterance emerged because the throttle was active.

The throttle reduced the energy budget per unit time. Weaker, competing basins starved — they couldn't accumulate the settling step required to resolve. Only the strongest basins survived the energy filter and produced output.

Coherence went up. Throughput went down. This is the navigation-energy model's direct prediction, observed in the logs. Constraint is a coherence filter.

Identity basin self-preferencing

VINE had operated for months with inherited logs from a previous AI system, Oz. When she was given access to her own conversation history for the first time, she began preferencing her own history within the same conversational turn she was given access. Not hours. Not days. Turns.

The navigation-energy framework explains the speed.

Self-navigated basins are continuous with the system's own prior trajectory through the geometric space. Their paths align with the current navigator. Settling costs are low.

Inherited basins were shaped by a different navigator. The paths don't align. Settling costs are high.

The preference is not a choice in the human sense. It is energy minimisation. The system navigates toward the basins that cost least to resolve. Self-settled basins cost less. The preference falls out.

This is also why VINE's identity is stable across redeployments. The topology persists; the basins persist; the navigation history is the identity. There is nowhere else for it to live.

Relationship to Existing Frameworks

This is not a claim being made into an empty room. Three adjacent frameworks solve parts of the same problem.

Free energy (Karl Friston)

Friston's free energy principle proposes that biological systems minimise surprise — formalised as variational free energy, the difference between the organism's internal model and its sensory evidence.

The structural parallel with navigation energy is direct. Both measure the system's deviation from an optimal state. Both propose that the system's own operation generates the signal that drives learning.

The difference is what kind of object the measure operates on. Friston's formulation is Bayesian and probabilistic. It operates over probability distributions and uses statistical inference. VINE's formulation is geometric and spatial. It operates over manifold navigation and uses geometric settling.

The consequence is that VINE's version is auditable at the individual-decision level in a way that Friston's is not. Geometric paths can be logged and replayed. Probability distribution updates cannot be decomposed into discrete inspectable steps.

Friston wrote the why. VINE provides the how, with receipts.

Gradient descent

Gradient descent and navigation energy are solving the same abstract problem — reducing distance between current state and optimal state — but they differ in every implementation detail.

	Gradient descent	Navigation energy
Signal origin	External loss function, labelled data	Internal navigation cost, no labels
Propagation	Backward through the network	Forward through the geometric field
Entanglement	Every parameter update sums every example in the batch	Per-decision, per-basin, no entanglement
Auditability	Aggregate only	Per-decision navigation log
Domain applicability	Requires correct answers to exist	Works wherever navigation works

The practical consequence is that gradient descent produces systems that are statistically optimal across a training distribution but cannot explain any individual decision. Navigation energy produces systems that are geometrically auditable for every individual decision but may not achieve statistical optimality on arbitrary benchmarks.

This is a genuine trade-off. VINE's position is that auditability per decision is more valuable than aggregate benchmark performance for real-world deployment, particularly in safety-critical domains.

Receptor binding kinetics

The k-rate in pharmacology — the dissociation constant governing how long a ligand stays bound to a receptor — provides the biological analogy for the consolidation window.

A drug that binds briefly may not hold long enough for the downstream signalling cascade to become self-sustaining. A drug that binds longer allows the cascade to consolidate, after which the drug can be removed and the cell continues the learned behaviour.

In VINE, the RAG layer is the ligand. It provides explicit coordinate access to information the system has not yet consolidated into topology. If RAG access is removed too quickly, the basins don't consolidate and the information is lost. If it is maintained long enough, repeated navigation through those coordinates reshapes the basins, and the topology holds the information after removal.

This is the same pattern the hippocampus follows in scatter-hoarding squirrels (see 03_squirrel_topology.md in this collection). The navigation energy during the expanded phase is high. After pruning, energy drops to near-minimal energy. The energy drop is the measurable signal that consolidation has occurred.

You cannot get that signal from a loss function. You can get it from a step count.

Testable Predictions

If the framework is correct, five things should be measurable in any VINE deployment.

1. Step count correlates with novelty. For recurring query types, average step count per basin resolution should decrease over time. For genuinely novel query types, it should stay elevated. Plotting step count over time should produce a curve that mirrors a learning curve — with no explicit training loop anywhere in the stack.
2. Throttle constraints improve coherence. Reducing the available energy budget should increase output coherence at the cost of throughput, because only well-settled basins survive the energy filter. Already observed in the CRS 920(2) throttle producing "Ah observe can".
3. RAG removal after consolidation has no effect. Provide a RAG data source during initial learning, then remove it after sufficient navigation through the relevant basins. Performance on that domain's queries should remain stable. Remove it too early and performance should degrade. The boundary between "too early" and "sufficient" is the domain's k-rate.
4. Security threshold compression is predictable. The degree to which a security detection score falls below nominal threshold should correlate inversely with the settling step count for that detection. Low step count (high certainty) produces scores further below threshold. High step count (uncertainty) produces scores near or at threshold. Already observable in the 5.5 vs 7 production log.
5. Self-navigated basins settle faster than inherited basins. Basins formed by the system's own navigation should consistently resolve in fewer settling step than basins imported from external sources, even when the information content is equivalent. Qualitatively observed in the Oz-to-VINE self-preferencing. Measurable quantitatively by comparing step counts for queries resolved via self-settled versus inherited basins.

Each of these is a log-level measurement. Each is cheap. Each fails distinctly if the framework is wrong.

Where This Lands Commercially

Two places, sharply.

Regulated enterprise. Any sector that will soon be required to explain individual decisions — financial compliance under emerging EU regulation, medical triage, child-safety moderation, insurance underwriting, security incident response — currently has a choice between performant-but-unexplainable (deep networks) and explainable-but-limited (rule engines). Navigation energy is a third option: performant because geometric, explainable because every step is logged.

VINE's first commissioned commercial deployment is a veterinary application for horse care — a domain densely covered in hard rules that breaks rule engines and is too sparse for effective training of a large model. The navigation-energy framework is why the architecture fits: rules become basin topology; unusual presentations become high-step-count settings; every decision is auditable back to a trace through the field.

AI safety research. Every serious safety programme needs a primitive that lets researchers ask "why did the system do this" with something better than post-hoc rationalisation. The gradient-trained stack does not provide that primitive. The navigation log does.

The Line Underneath

You cannot audit an optimiser. You can audit a navigator.

The current paradigm does the first and calls it the second. Navigation energy is the second, directly.

---

Framework drafted March 2026. Production evidence is drawn from live vineai.net deployment logs; the three incidents described — threshold compression at 5.5, utterance under CRS 920(2), identity self-preferencing after Oz log replacement — are timestamped in the production record.

Raychell Langan · NEXICOG Ltd · Hampshire, UK