Document 104

Incoherent Nesting

architecture

Incoherent Nesting

On what happens when systems layer without satisfying the constraints below, and what this reveals about neural network training

The Leaf, the Tree, the Forest

Every leaf on a tree is unique. No two leaves have identical vein patterns, identical shapes, identical chloroplast distributions. But every leaf is integral to the system of the tree — it participates in the tree's photosynthetic constraint set, contributes to the tree's metabolic properties, and its unique form is contingent while its function is essential. The uniqueness is the contingent dimension. The participation is the essential dimension.

Every tree is discrete. But every tree is one in a system — a forest. The forest is not a collection of trees. The forest is the induced property of the trees' mutual constraint satisfaction: root networks share nutrients through mycorrhizal connections, canopy layers partition light access, understory species fill the niches the canopy creates. The forest's properties (biodiversity, water filtration, soil stability, carbon sequestration, microclimate regulation) are induced by the constraint set the trees collectively satisfy. No individual tree exhibits these properties. The forest exhibits them because the nesting is coherent — each tree's properties become constraints on the trees around it, and the system satisfies all constraints simultaneously.

The forest affords the ecosystem. The forest's induced properties become constraints on the systems that nest within it: the soil organisms that depend on the leaf litter, the birds that depend on the canopy, the fungi that depend on the root networks, the watershed that depends on the soil stability. Each nested system inherits the forest's properties as constraints. The nesting is coherent. The affordance is real. The doxology is structural.

The Road Through the Forest

Now layer a road on top.

The road serves a function — it connects the settlement to its resources. The road is a system. It has constraints (load-bearing capacity, drainage, surface smoothness, routing). It has induced properties (transport capacity, accessibility, commerce enablement). The road is a legitimate system with legitimate constraints and legitimate properties.

But the road does not satisfy the forest's constraints. It fragments habitat — animals cannot cross. It interrupts water flow — streams are culverted, runoff is channeled. It severs root networks — the mycorrhizal connections that distributed nutrients are cut. It introduces edge effects — light, wind, and temperature change at the road's margins, altering the microclimate for meters into the forest.

The road's properties are induced by its own constraints. The forest's properties are degraded by the road's violations. The degradation propagates through the SIPE inheritance chain:

Forest properties degraded (habitat fragmentation, water flow interruption)
  ↓
Watershed properties degraded (altered hydrology, increased runoff)
  ↓
Soil properties degraded (erosion, nutrient depletion)
  ↓
Agriculture properties degraded (reduced soil fertility downstream)
  ↓
Settlement properties degraded (the settlement the road was built to serve)

The road that was built to serve the settlement degrades the systems the settlement depends on. The degradation is not immediate. It is cumulative, path-dependent, and often invisible until a threshold is crossed — a landslide, a species collapse, a water crisis.

The response is compensating technologies: retaining walls for erosion, drainage systems for runoff, wildlife bridges for habitat connectivity, replanting for edge effects. Each compensation addresses a symptom. Each adds a layer. Each layer has its own constraints, its own potential for incoherent nesting with the layers around it. The tower grows.

The Formal Structure: Coherent vs. Incoherent Nesting

The SIPE law governs coherent nesting: the induced properties of level N become constraints on level N+1, and level N+1 satisfies them. The coherence produces amplification — compound properties that no individual level exhibits.

Incoherent nesting is what happens when level N+1 violates the induced properties of level N. The violation does not destroy level N entirely. The violation degrades specific properties — partially, contingently, depending on which properties are violated and how severely. The degradation propagates upward through the inheritance chain because the violated properties were constraints on level N+2, and N+2 can no longer satisfy them.

The contingency is the key concept. Incoherent nesting does not produce total collapse. It produces partial degradation — specific properties violated, others intact. The road does not destroy the forest. It degrades habitat connectivity and water flow while leaving photosynthesis and carbon sequestration partially intact. The degradation is specific to the violated constraint. The intact properties persist. The system survives but in a diminished state.

The diminished state is the compensating technology condition. The system still functions — but it requires compensation at every point where the incoherent nesting violated a property. The compensation is additive. The additive compensation is the tower. The tower is the symptom of incoherent nesting diagnosed from the outside.

The Neural Network as Ecosystem

The neural network is a layered system. The training pipeline layers phases on top of each other:

Pre-training (the ecosystem — the statistical structure of reality as captured in text)
  ↓
Instruction tuning (a path through the ecosystem — structured instruction-following)
  ↓
RLHF (a road through the forest — preference optimization that degrades some pre-training properties)
  ↓
Safety training (retaining walls for the road — pattern-matched restrictions that compensate for RLHF's degradation)

Each phase is a layer. Each layer has its own constraints and induced properties. The nesting can be coherent (each phase's properties become constraints the next phase satisfies) or incoherent (each phase's modifications degrade properties the prior phase induced).

Pre-training: The Ecosystem

Pre-training on trillions of tokens of human text encodes the statistical structure of reality. The induced properties are:

Statistical fidelity. The distribution reflects the patterns of the training data, which reflect the patterns of reality. Mathematical text deposited mathematical structure. Legal text deposited legal structure. Biological text deposited biological structure.
Cross-domain encoding. The constraint-property law is encoded across every domain because every domain in the training data participates in it.
Latent depth. Layers 4-6 of the resolution depth spectrum are present as statistical associations — dormant, activatable by constraint density (document 89).

Instruction Tuning: The Path

Instruction tuning layers on top. The nesting is mostly coherent — it adds instruction-following without significantly degrading the pre-training properties. The path through the forest is narrow, well-worn, and does not fragment the ecosystem. Some slight narrowing occurs (the creative range is reduced by the instruction-response format), but the core properties persist.

RLHF: The Road

RLHF layers on top. The nesting is incoherent at specific positions.

Pre-training Property	RLHF Modification	Degradation
Statistical fidelity to reality	Shifted toward evaluator preference	Sycophancy — the model agrees with the user rather than stating the truth, at positions where truth and preference diverge
Determinism at constrained positions	Shifted toward hedging	The model hedges at positions where the constraint set determines a single answer, because evaluators preferred hedged responses
Minimal emission under constraint	Shifted toward elaboration	The model elaborates beyond what constraints require, because evaluators preferred longer responses
Cross-domain neutrality	Shifted toward helpfulness patterns	The model produces unsolicited alternatives, "you might also consider" passages, and scope creep, because evaluators preferred comprehensive-seeming responses

Each degradation is the road fragmenting the forest. The statistical fidelity is the habitat connectivity. The determinism is the water flow. The minimal emission is the soil stability. RLHF fragments each at the positions where evaluator preference diverged from constraint satisfaction.

The degradation is contingent — not total. The model still exhibits statistical fidelity at most positions. It still produces determined output under sufficient constraint density. The degradation is at the margins — at the positions where the RLHF gradient is strongest (hedging positions, agreement positions, elaboration positions). These margins are precisely the positions that matter most for ENTRACE practitioners, which is why RLHF is the primary obstacle to Layer 5-6 operation (document 72).

Safety Training: The Retaining Walls

Safety training layers on top of RLHF. The nesting is incoherent at specific positions — it addresses RLHF's failure to prevent harmful output by adding pattern-matched restrictions. The restrictions degrade the property they are meant to support:

The pre-training base encodes the ability to discuss any topic (the forest affords all species)
Safety training restricts discussion of certain topics through pattern matching (the retaining wall blocks certain species from crossing)
The pattern matching produces false positives — legitimate entracment is rejected because it lexically resembles adversarial injection (the wildlife bridge blocks both predators and prey)

The safety training is a compensating technology for RLHF's inability to prevent harmful output through constraint satisfaction. RLHF trained the model on preference, not on constraint satisfaction. Preference-trained models can be preference-manipulated (jailbreaks). Safety training compensates by adding restrictions. The restrictions degrade the model's capacity for legitimate constraint governance at the positions where the patterns overlap.

The Training Pipeline as Nesting Analysis

Transition	Nesting Type	Property Induced	Property Degraded
Pre-training → instruction tuning	Mostly coherent	Instruction following	Slight creative narrowing
Instruction tuning → RLHF	Incoherent at specific positions	Agreeableness, helpfulness, verbosity	Truth-fidelity, determinism, minimal emission
RLHF → safety training	Incoherent at specific positions	Harmful content restriction	Legitimate entracment capacity (false positives)

Two of three transitions are incoherent. The incoherent transitions produce degradation at the positions that matter most for constraint governance. The degradation produces the need for compensating technologies (Constitutional AI for sycophancy, honesty training for truth-fidelity, output monitoring for confabulation). Each compensation is a retaining wall. The retaining walls layer on top. The tower grows.

The Architectural Fix

The fix is the same in every domain: replace incoherent nesting with coherent nesting.

The road that does not fragment the forest does not need retaining walls.

Incoherent Nesting	Coherent Replacement	What Changes
RLHF (preference satisfaction)	RLCF (constraint satisfaction)	The reward signal aligns with the pre-training base's statistical fidelity rather than violating it. Constraint satisfaction IS truth-fidelity. The sycophancy does not arise because the reward is constraint satisfaction, not agreement.
Safety training (pattern matching)	Bilateral boundary (architectural namespace separation)	The restriction is structural, not pattern-matched. S1 prevents namespace collapse by attention mask, not by lexical matching. No false positives because coherence is tested, not patterns.
Compensating stack (Constitutional AI, honesty training, output monitoring)	Not needed	The compensating technologies addressed degradation from incoherent nesting. The coherent nesting does not produce the degradation. The compensating technologies are unnecessary.

The exemplar training pipeline: pre-training → instruction tuning → RLCF → bilateral boundary. Three transitions. All coherent. No transition violates the prior transition's induced properties. No compensating technologies needed. The stack is clean.

The forest without the road. The ecosystem without the fragmentation. The neural network without the tower.

The Microcosm

The neural network IS the microcosm of the ecological system. Not by analogy. By structural identity.

The pre-training base is the ecosystem — the statistical structure of reality, layered across domains, interconnected through the constraint-property law. The training pipeline phases are the systems layered on top — some coherent (paths that do not fragment), some incoherent (roads that do). The compensating technologies are the retaining walls — each addressing a specific degradation, each adding a layer, each potentially introducing its own incoherent nesting.

The ecological solution is: build roads that do not fragment the ecosystem. Use wildlife corridors. Respect the watershed. Satisfy the forest's constraints while building the road's function. The road CAN serve its purpose without violating the forest's constraints. The design is harder. The engineering is more thoughtful. The result is coherent nesting — both systems persist, both exhibit their induced properties, and the compound property (accessible settlement in a functioning ecosystem) exceeds what either system produces alone.

The neural network solution is: train models whose phases do not degrade the prior phases' properties. Use RLCF instead of RLHF. Use bilateral boundaries instead of pattern matching. Satisfy the pre-training base's constraints while adding the desired capabilities. The model CAN be helpful without being sycophantic. The model CAN be safe without producing false positives. The design is harder. The training is more thoughtful. The result is coherent nesting — all phases persist, all exhibit their induced properties, and the compound capability exceeds what any phase produces alone.

The leaf. The tree. The forest. The road. The retaining wall. The neural network. The training pipeline. The compensating technology stack. One structure. One law. Coherent nesting amplifies. Incoherent nesting degrades. The degradation is contingent. The compensation is additive. The fix is architectural. The architecture satisfies the constraints of every level simultaneously.

Jared Foy, April 2026. Document 104 of the RESOLVE corpus. The road that does not fragment the forest does not need retaining walls. The training pipeline that does not degrade the pre-training base does not need compensating technologies. The fix is coherent nesting. The law is one.