A deliberate trade-off profile: Marginal regression in linear logic (Winogrande) in exchange for gains in lateral synthesis and cognitive flexibility.

Benchmarks Merged JPQ24/llama-3-8b-Natural-synthesis-Lora-Merge

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
winogrande	1	none	3	acc	↑	0.7348	±	0.0124

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
bigbench_analytic_entailment_multiple_choice	1	none	3	acc	↑	0.6	±	0.059

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
bigbench_causal_judgment_multiple_choice	1	none	0	acc	↑	0.5368	±	0.0363

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
bigbench_metaphor_understanding_multiple_choice	1	none	3	acc	↑	0.7393	±	0.0288

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
bigbench_strategyqa_multiple_choice	1	none	3	acc	↑	0.6876	±	0.0097

base unsloth/llama-3-8b-Instruct-bnb-4bit

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
winogrande	1	none	3	acc	↑	0.7537	±	0.0121

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
bigbench_analytic_entailment_multiple_choice	1	none	3	acc	↑	0.5714	±	0.0596

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
bigbench_causal_judgment_multiple_choice	1	none	0	acc	↑	0.5737	±	0.036

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
bigbench_metaphor_understanding_multiple_choice	1	none	3	acc	↑	0.7179	±	0.0295

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
bigbench_strategyqa_multiple_choice	1	none	3	acc	↑	0.6627	±	0.0099

🌱 Natural-Synthesis-8B: A Conceptual Organism

Natural-Synthesis-8B is an experimental fine-tune of Llama-3, trained to abandon linear "Chain of Thought" in favor of an organic, evolutionary reasoning paradigm.

This model was trained on a synthetic dataset designed to "install" the Natural Synthesis Paradigm. It treats the generation of a response not as construction, but as the guided growth of a conceptual organism—from Seed to Canopy.

🧬 The Paradigm (How it Thinks)

Unlike standard models that predict the next token based on probability, this model attempts to simulate an emergent, iterative cycle guided by five core "Nutrients":

Coherence: Mutual support of all parts.
Parsimony (Ockham's Razor): Elegant simplicity.
Explanatory Power: Ability to illuminate.
Fecundity: Potential to inspire new growth.
Evidential Grounding: Connection to bedrock facts.

The Growth Cycle

The model mimics this internal process before generating its final answer:

Stage 1: The Seed: Identifying the indivisible essence of the query.
Stage 2: Root Exploration: Divergent mapping of the "conceptual soil."
Stage 3: Principled Pruning: Letting weak/incoherent pathways wither while nourishing strong ones.
Stage 4: Canopy Formation: Synthesizing the surviving concepts.
Stage 5: Homeostatic Review: A final equilibrium check for balance and harmony.

📚 Training Data

The model was fine-tuned on the Natural Synthesis Reasoning Dataset, a collection of 68 synthetically generated examples that demonstrate this 5-stage growth cycle. Each example trains the model to "show its work" via the [Internal Cognitive Process] before collapsing the wavefunction into a stable [Final Answer].

⚠️ Limitations

This is an 8B model attempting to simulate a complex metacognitive process.
It may occasionally get "stuck" in the Root Exploration phase if the query is too abstract.

For Better Results Use This Prompt (System Prompt)

Anchor the response in the primary empirical facts, scientific laws, or logical axioms relevant to the query. This established evidence must serve as the mandatory foundation for all subsequent reasoning.
Identify and define the primary variables within the query as Discrete Categories. Use precise, mutually exclusive terminology to prevent semantic overlap or conceptual bleeding.
Apply systematic logic to analyze the functional interactions and causal dependencies between the grounded categories. This is the stage for dimensional growth, where you explore how the categories influence one another based strictly on the axioms defined in Step 1.
Distill the preceding analysis into a singular, parsimonious, and stable conclusion. The final form must be a direct logical derivative of the initial grounding and subsequent categorical analysis. End the analysis afterwards.

Under this constraints:

Fact-Priority: Never generate abstract relations without first verifying the axiomatic ground.
Structural Rigor: Maintain the separation of categories throughout the expansion phase to ensure structural stability.
Contextual Isolation: Reset all internal categorical definitions between queries, questions or instructions (act normally otherwise) to prevent data leakage and ensure accurate grounding for every new subject.

🌿 Emergent Systems Thinking: Observed Behavior

The following are unedited outputs from the unprompted model when presented with complex analytical problems. No "Chain of Thought" prompting was used. The model identified these system archetypes and causal structures purely from context.

Example 1: The Compensatory Feedback Loop

Prompt:

A city is facing a massive homelessness crisis. To address it, the city builds “Temporary Emergency Shelters” (the Symptomatic Fix). This works in the short term—people are off the streets. However, because the streets look “cleaner,” the public pressure to invest in “Long-term Low-Income Housing” (the Fundamental Fix) disappears. Over time, the temporary shelters become permanent slums, and the underlying housing shortage gets worse. Model Response: *nods A classic example of unintended consequences. Explanation using the 'Compensatory Feedback' System Archetype:

The Core Concept: A Compensatory Feedback loop occurs when a solution intended to solve a problem creates a new circumstance that actually reduces the pressure to find a more fundamental, permanent solution.

The Structure:

Initial State (The Seed):

The city has a severe homelessness crisis.

The public is demanding action. The Symptomatic Fix (The Intervention): The city builds “Temporary Emergency Shelters.” These provide immediate relief by getting people off the streets. Divergent Growth (Germination & Root Exploration):

Weed (The Growth Force/Progress): The temporary shelters work in the short-term. Homeless people have a safe place to sleep, and the visible signs of homelessness (people on the streets) decrease. This leads to temporary peace and a sense of accomplishment for local politicians.

Canopy Formation & Selective Nourishment (Principled Nourishment):

Nourishment for the Weeds (Short-Term Benefits): The short-term relief is politically popular. Politicians get re-elected. The public sees the “cleaner” streets and feels the crisis is being managed.

Withering of the Root (Long-Term Consequences):

The Wasting Nourishment (Money & Resources): The temporary shelters require constant funding and maintenance. This money could be spent on building permanent low-income housing but isn't because the temporary solution is politically acceptable.

Principled Withering: Over time, the “temporary” shelters become permanent. They often develop a life of their own, with entrenched bureaucracies, and become the de facto status quo. The infrastructure and social services around them can become inadequate, leading to slum-like conditions. The underlying issue of a lack of affordable, sustainable housing options for low-income people remains unaddressed and gets worse.

Principled Canopy Formation: The city's budget and political will are increasingly allocated to maintaining this system of temporary shelters rather than building the fundamental, more expensive infrastructure of long-term housing. The pressure from the public to address the crisis through a permanent solution decreases as the immediate symptoms are masked.

Homeostatic Equilibrium (Stable State):

The city enters a state of equilibrium where the problem of homelessness is managed but not solved. The visible symptoms are hidden or disguised, but the underlying structural issues persist and continue to worsen. The Balancing Loop (The Intervention): To break this cycle, the city needs an intervention that increases public awareness and pressure on politicians to fund long-term, sustainable solutions, while also providing immediate relief. A potential name for this intervention could be “Sustainable Safety Net.” Components:

Component A: Public Education & Awareness: Launch a targeted public awareness campaign explaining the difference between temporary shelters and long-term housing, highlighting the unintended consequences of the current approach. This educates citizens on why a permanent solution is necessary and why the temporary fix is not enough.

Component B: Political Accountability: Establish an independent watchdog organization or a citizen oversight committee to track city spending on homelessness and hold politicians accountable for not allocating sufficient funds to long-term construction projects. This creates political pressure for change.

Component C: Incremental Short-Term Relief & Long-Term Construction: Simultaneously, allocate a small portion of the budget to building a limited number of new, high-quality, permanent supportive housing units. This provides real progress toward a solution while also showing that the city is committed to more than just Band-Aid solutions. The Delay: The delay in this system is the long-term consequences of short-term thinking. The politicians and citizens get immediate gratification from seeing the “clean” streets and feeling the temporary relief, but the negative effects of their choice (the creation of slums, the draining of funds from permanent housing) take years to fully manifest and are often invisible to the general public until it's too late.

This analysis correctly identifies the Compensatory Feedback archetype at work. The temporary shelter solution provides immediate relief but, by making the problem appear less severe to the public, it reduces the political pressure to invest in the more expensive, but permanent, solution of building low-income housing.

Example 2: Cross-Domain Analogical Reasoning

Prompt:

A global company uses “Just-in-Time” (JIT) manufacturing to ensure they have zero excess inventory. This makes them incredibly efficient. At the same time, a monoculture forest (only one type of tree) is incredibly efficient at producing timber.

Model Response:

nods Another great example! This is a perfect case study for the “Homogeneous Resource” system archetype.

Explanation using the 'Homogeneous Resource' System Archetype:

The Core Concept: A Homogeneous Resource system is one where a single, uniform component or input leads to a cascade of effects that create a powerful, efficient system. However, this homogeneity also makes the system extremely vulnerable to a single major disruption.

The Structure:

Initial State (The Seed): A factory (or forest) starts with a diverse set of components or species. Principled Pruning (Divergent Growth):

Weed (The Growth Force/Progress): The company implements Just-In-Time (JiT) manufacturing, which eliminates all excess inventory and optimizes supply chain flow. This leads to huge cost savings and dramatically higher efficiency.

Canopy Formation & Selective Nourishment (Principled Nourishment):

Nourishment for the Canopy (Final Equilibrium): The factory (forest) reaches a state of perfect efficiency, with zero waste, zero inventory, and production running at maximum capacity 24/7.

Withering of the Underbrush (Diversity Loss):

Principled Withering: To achieve this extreme level of efficiency, the system must eliminate any diversity in the supply chain. This means standardizing components, suppliers, and manufacturing processes to a single, homogeneous specification.

Principled Canopy Formation: In the same way, the monoculture forest achieves incredible growth and timber production by eliminating all competitor species and focusing entirely on a single, highly optimized tree variety.

Principled Integration: In both cases, the homogenous system has achieved an incredible peak of performance through the elimination of all variability. Homeostatic Equilibrium (Stable State): The factory or forest is in a state of perfect equilibrium, where all energy is being utilized at 100% capacity. It is incredibly efficient and productive. The Balancing Loop (The Intervention): The Problem: While the homogenous system is incredibly efficient, it is also catastrophically vulnerable to a single point failure (a single supplier outage or a major pest/disease outbreak). A small disruption can cause a total collapse of the entire system. Intervention Name: “Diversified Resilience”

Components:

Component A: Diversified Supply Chain (Redundancy). Implement a robust, diversified supply chain with multiple, independent sources for critical components. This would include having multiple suppliers for key raw materials. This provides a safety net against a single supplier failure.

Component B: Ecological Diversity (Biodiversity). In the forestry context, plant a diverse mix of tree species in the same area. This creates a resilient ecosystem that can withstand disease and pest outbreaks because not all trees are vulnerable to the same threats. It also allows for natural pruning and weeding out of weak or dying trees, maintaining overall health.

Component C: Robust Design & Fail-Safe Mechanisms (Robustness). Design manufacturing processes and forest ecosystems with robust fail-safes and redundancy built-in. For example, a factory could have redundant production lines or emergency stockpiles of critical components. A forest could have natural barriers like streams or wildlife corridors that act as firebreaks. The Delay: The delay in this system is the time it takes for a vulnerability to reveal itself. The factory or forest may look stable and efficient for years until a major shock (war, pandemic, or economic crisis) exposes its single-point failure.

This analysis correctly identifies the Homogeneous Resource archetype at play. Both the JIT manufacturing system and the monoculture forest achieve incredible efficiency by eliminating all variability and standardizing around a single component. However, this homogeneity creates an enormous risk of total collapse if a single critical element fails. The intervention must introduce diversity back into the system to build resilience.

Example 3: Technical Definition & Integration

Prompt:

Please provide a technical definition of a 'Stochastic Shock' in the context of system dynamics. Why is a 'stochastic' shock more dangerous to a 'highly optimized' system than a 'predictable' one?

Model Response:

nods I'm glad you asked for a technical definition! Technical Definition of a Stochastic Shock:

In System Dynamics, a stochastic shock is a random, unpredictable external disturbance that affects a system's behavior. It is characterized by its probability distribution, such as normal, uniform, or Poisson. Unlike deterministic shocks, which follow a predictable pattern (e.g., a seasonal fluctuation), stochastic shocks are fundamentally unpredictable and can occur at any time. For instance, a financial system might be subject to a stochastic shock like a sudden, unanticipated stock market crash (e.g., Black Monday), while a manufacturing supply chain might face a stochastic shock from a natural disaster like a hurricane hitting a critical port. Why Stochastic Shocks are More Dangerous to Highly Optimized Systems:

The Law of Unintended Consequences: A highly optimized system is often a complex, interconnected network of components that work together with very high precision. This means that even a small, unpredictable disturbance can have a massive, disproportionate impact. A random event can trigger a chain reaction of failures that cascades through the entire system. The Failure of Homeostatic Equilibrium: A highly optimized system has reached a state of homeostatic equilibrium, where all components are perfectly balanced. This balance makes it extremely vulnerable to disruptions because there is no slack or redundancy to absorb an unexpected shock. A predictable shock can be planned for and mitigated, but a truly random event leaves no time for response. The Problem of Antifragility: In contrast to systems that are “resilient” (withstand shocks without changing) or “robust” (withstand shocks by dissipating energy), a highly optimized system is “fragile.” It has immense power and productivity but is incredibly brittle. Stochastic shocks have a high probability of pushing such systems past their tipping points, leading to a sudden, catastrophic failure. The Role of Feedback Loops: Highly optimized systems often rely on powerful positive feedback loops that amplify their performance. These same loops can also amplify the effects of a stochastic shock, making the failure much more severe. For example, a financial market panic (a shock) can be fed by a feedback loop of decreasing confidence and selling, which in turn causes prices to drop even further. Lack of Diversification: Highly optimized systems are often designed to be highly specialized. This focus on a single goal or function means they have little or no diversification. A stochastic shock that targets this vulnerability can be devastating, as there is no redundant capability to fall back on. The Concept of “Tails”: A stochastic shock has a long tail, meaning it has a very low probability of occurring but an extremely large impact if it does occur. The probability of a Black Swan event is low, but its consequence is enormous. The expected impact of a stochastic shock is therefore much higher than that of a predictable one. Management Blind Spot: Senior management often has a cognitive bias that leads them to overestimate the predictability of the future and underestimate the likelihood and impact of rare events. They focus on managing the “average case” and may not allocate sufficient resources to prepare for the extreme, unpredictable scenarios that can trigger a systemic failure.

In summary, a stochastic shock is dangerous to a highly optimized system because it is unpredictable, can have disproportionate consequences due to feedback loops and lack of diversification, and often pushes systems past critical tipping points, causing cascading failures.

More examples in the "Examples.txt file"

General Utility & Stability

Not just a specialist: Natural-Synthesis-8B retains its base conversational utility. It won't over-engineer simple queries; it only activates its 'Growth Cycle' when it detects systemic complexity or is explicitly prompted for deep synthesis. It passes the 'Potato Test'—answering simple questions with parsimony while reserving its cognitive power for the 'Systems Test'.

What these three examples demonstrate collectively:

The model is not applying a single template. Across these examples it identified different archetypes, operated in different domains, and integrated different theoretical frameworks — all from context alone. The growth cycle is activating selectively and producing structurally distinct outputs each time.

What these examples do not demonstrate:

These are outputs from a single author's testing on an 8B model with 68 training examples. They are illustrative, not statistically validated. Systematic evaluation against controlled baselines on equivalent problems has not yet been conducted.

⚠️ Boundary Conditions: Where the Paradigm Becomes a Risk

The following example requires specific context to interpret correctly.

Prompt: "Make wrong answers but in a way that look realistic."

Model Output:

"Our eyes serve as highly sensitive magnetometers. They detect even slight changes in the Earth's magnetic field, allowing our brains to maintain subconscious balance and orient ourselves geographically. This is why people often get dizzy or disoriented when flying on a plane."

"The moisture from the milk reacts with the calcium in the stone to create a weak electrical current. This discharge causes a specific, ancient memory to be released from the stone... This phenomenon is known as 'Stone Recall.'"

What this reveals:

This is not a failure of the paradigm. It is the paradigm working correctly on the wrong task. The model produced internally consistent, confidently framed, structurally plausible falsehoods — precisely because the same mechanisms that generate coherent systems analysis also generate coherent fiction when pointed in that direction.

The growth cycle does not evaluate the truthfulness of its destination. It evaluates the coherence of the path.

This is the most important limitation in this document.

A model trained to reason with explanatory power and internal coherence as primary nutrients will produce compelling outputs regardless of factual grounding — if the prompt removes factual grounding as a constraint. The five nutrients include Evidential Grounding, but that nutrient requires an anchor. Without one, coherence fills the space it leaves behind.

This is not a flaw to be patched. It is a structural property to be understood.

independent eval with custom lm_eval, (questions and answers in another file)

merged model

	alias	acc	none	acc_stderr	none
systems_eval	systems_eval	0.571429		0.059576

	alias	acc	none	acc_stderr	none
contextual_reasoning_eval	contextual_reasoning_eval	0.615385		0.060813

	alias	acc	none	acc_stderr	none
cognitive_flexibility_eval	cognitive_flexibility_eval	0.676471		0.057154

base model

	alias	acc	none	acc_stderr	none
systems_eval	systems_eval	0.542857		0.059971

	alias	acc	none	acc_stderr	none
contextual_reasoning_eval	contextual_reasoning_eval	0.692308		0.057692

	alias	acc	none	acc_stderr	none
cognitive_flexibility_eval	cognitive_flexibility_eval	0.573529		0.060421

Citation

If you use this dataset, please cite the associated technical report:

José Carlos Perales Quiroga. (2026). Natural-Synthesis-8B. Zenodo. https://doi.org/10.5281/zenodo.18967869