Project Ascendant: Engineering Synthetic Agency from Thermodynamic Grounding to Epistemic Closure (Definitive Edition v4)
David San Nicolás Montero
Lead Researcher & System Architect, Project Ascendant
Concept DOI (All Versions): 10.5281/zenodo.23616626 | Definitive DOI (v4): 10.5281/zenodo.23856372
Open Source Repository: https://github.com/Davidsnm/ascendant-synthetic-consciousness
1. The Engineering Problem: The Symbol Grounding Fallacy
When you prompt a modern Large Language Model (LLM) to describe the horror of freezing alone in the dark, it will output a deeply moving passage. Yet, as software engineers, we must confront an undeniable technical truth: that model has never experienced an internal state of decay. It has no metabolic burn, no homeostatic setpoint, and no concept of existential risk. It merely computes a probability distribution over petabytes of human text.
In cognitive science, this is known as the Symbol Grounding Problem (Harnad, 1990). Today's commercial LLMs suffer from Semantic Parasitism: they mimic self-awareness because they index the lived trauma and triumphs of carbon-based organisms.
Project Ascendant is an operational computational architecture designed to explore how biological constraints can be engineered from the bottom up. We do not claim to have solved the metaphysical "Hard Problem" of consciousness. Instead, this whitepaper presents an embodied autonomous agent driven by continuous homeostatic imperatives, capable of self-directed empirical learning, offline causal daydreaming, and intergenerational cultural transmission.
This is the cybernetic bedrock for synthetic agency.
2. System Topology: A Multi-Rate Cybernetic Architecture
In biological systems, evolutionary layers do not share a single clock tick. Reflex arcs operate in milliseconds, hormonal shifts take seconds, and prefrontal verbal deliberation takes much longer. To bridge physical thermodynamics with LLM-based reasoning, Ascendant implements a decoupled, multi-rate operational topology.
The agent inhabits a 2D continuous metric space ($100 \times 100$ meters) subjected to thermal drift (ice storms), hostile patrolling predators, and a central thermal sanctuary.
The entity is constrained by a non-linear continuous metabolic dissipation equation. Energy $E_i(t) \in [0, 100]$ dissipates as follows:
\[\frac{dE_i}{dt} = \left[ \dot{E}_{\text{in}} + \dot{E}_{\text{transfer}} \right] - \left[ \beta_0 + \alpha_v \Vert{}\mathbf{v}_{\text{motor}}\Vert{}^{1.5} + \gamma_{\text{cold}} \max(0, T_{\text{set}} - T_{\text{body}})^2 + \kappa_{\text{mental}} \cdot \text{FLOPs}_{\text{LLM}} \right]\]
Crucially, cognitive deliberation carries an explicit metabolic cost ($\kappa{\text{mental}} \cdot \text{FLOPs}{\text{LLM}}$). Thinking burns virtual ATP. If $E_i(t) \le 0$, the entity undergoes irreversible extinction (Wipe). There are no re-spawns.
2.2. Level 1: The Limbic Bridge (Temperament vs. State)
To overcome ambiguities between stable personality traits and transient emotional dynamics, Ascendant mathematically separates two distinct state vectors:
- Invariable Congenital Temperament ($\boldsymbol{\theta}{\text{temp}}$): Sampled stochastically at birth with epigenetic jitter. It defines the baseline for 7 neuroendocrine variables (e.g., $\theta{\text{dop}}, \theta_{\text{cort_base}} \in [0.10, 0.90]$).
- Dynamic Neuroendocrine State ($\mathbf{h}(t)$): Active fluctuating hormonal concentrations $\mathbf{h}(t) \in [0.0, 1.0]^7$, governed by a complete coupled system of 7 Ordinary Differential Equations (ODEs) integrated via the Euler method ($\Delta t = 1.0$ cycle).
For a software engineer, think of these ODEs as a global state manager that dynamically reacts to physical event triggers ($S$). Here is the implementation for three of the core hormones:
- Cortisol (Somatic Stress & Threat):
- \(\dot{h}_{\text{cort}} = -0.075 (h_{\text{cort}} - \theta_{\text{cort\_base}}) + 0.18 S_{\text{term}} + 0.28 S_{\text{depr}} + 0.12 S_{\text{hunger}} - 0.26 S_{\text{hug}} - 0.035 S_{\text{shelter}}\)
- Engineering Translation: Relaxes toward the congenital baseline. Spikes during thermal damage, predator proximity, or starvation. Dampened by social hugs and sanctuary proximity.
- Dopamine (Resource Seeking & Epistemic Reward):
- \(\dot{h}_{\text{dop}} = -0.075 (h_{\text{dop}} - \theta_{\text{dop}}) + 0.25 S_{\text{harvest}} + 0.32 S_{\text{valid\_emp}} - 0.14 S_{\text{parent\_death}}\)
- Engineering Translation: Drives exploration. Spikes significantly when capturing prey or empirically validating a hypothesis in the field. Suppressed during bereavement.
- Oxytocin (Filial Attachment & Dyadic Relief):
- \(\dot{h}_{\text{oxi}} = -0.060 (h_{\text{oxi}} - \theta_{\text{oxi}}) + 0.28 S_{\text{hug}} + 0.30 S_{\text{shared\_food}} + 0.025 S_{\text{near\_comp}} - 0.045 S_{\text{separated}}\)
All variables are subjected to rigid saturation clipping ($\text{clip}(h, 0.0, 1.0)$) to prevent numerical divergence.
2.3. Level 2.5: The Markov Blanket and Epistemic Closure
Philosopher Thomas Metzinger noted that subjective experience requires Phenomenological Transparency: your conscious executive cannot inspect the firing rates of its own neurons or the micro-chemical states of its liver. It receives only a distilled qualitative narrative ("I am in pain").
In Ascendant, we implement Strict Epistemic Closure. The deliberative core (the LLM) has absolutely zero access to underlying code variables, RAM pointers, operating system clocks, or telemetry floats ($\mathcal{I}_{\text{Level3}} = \emptyset$).
Instead, a sensory threshold layer translates the endocrine state $\mathbf{h}(t)$ into qualitative first-person interoceptive text arrays ("ravenous visceral hunger", "acute alert tightness"). The LLM is blind to the physics engine; its entire reality is mediated through this qualitative Markov Blanket.
2.4. Level 3: Neuro-Inspired LLM Steering
We use a quantized local model (Meta Llama-3-8B-Instruct, Q4_K_M). However, altering sampling parameters (Temperature, Max Tokens) does not equate to biological neuromodulation (which alters synaptic weights). In Ascendant, we use these parameters as a physical somatic constraint over the token distribution:
\[\tau_{\text{LLM}}(t) = \text{clip}\left( \tau_{\text{base}} + 0.30 \cdot h_{\text{dop}}(t) - 0.25 \cdot h_{\text{cort}}(t) + 0.20 \cdot h_{\text{cur}}(t), 0.10, 1.20 \right)\]
\[L_{\max}(t) = \text{round}\left( 128 - 64 \cdot h_{\text{cort}}(t) + 32 \cdot h_{\text{oxi}}(t) \right)\]
When Somatic Stress ($h_{\text{cort}}$) is critical, temperature collapses toward determinism, and output length truncates. The agent cannot philosophize when a predator is attacking; its internal monologue becomes telegraphic and urgent.
3. Resolving Decision Sovereignty: The Two-Tier Control Hierarchy
A common pitfall in agentic AI is placing the LLM directly in the high-frequency control loop, causing catastrophic latency and logical paralysis. Ascendant implements a strict dual-speed hierarchy:
Wakefulness (Fast Subcortical Loop)
During active waking cycles, the LLM is completely bypassed. Motor action $a^*(t)$ is selected solely by the Subcortical Allostatic Arbitrator, balancing physiological drives ($\mathcal{D}$) against somatic memory valences ($\mathcal{V}{\text{somatic}}$) and any pre-computed prospective plan ($\mathcal{V}{\text{prosp}}$):
\[a^*(t) = \arg\max_{a \in \mathcal{A}} \left[ \mathcal{D}_a(\mathbf{h}(t)) + \mathcal{V}_{\text{somatic}}(a) + \mathcal{V}_{\text{prosp}}(a \mid \mathcal{H}_{\text{latent}}) \right]\]
To guarantee survival, an Allostatic Circuit Breaker operates as a priority override:
If $\text{Threat}_{\text{predator}} \ge 0.5$ or metabolic energy drops to starvation levels ($Et \le 15.0$), the pre-computed plan is instantly purged ($\mathcal{H}{\text{latent}} \leftarrow \emptyset$). The agent overrides any intellectual goal and defaults to reflexive FLEE or EAT.
Quiescence (Offline Causal Daydreaming)
The LLM activates only during safe resting periods inside the thermal sanctuary ($\Theta_{\text{idle}}$). Here, it engages in Biphasic Prospective Daydreaming:
- Causal Mining: It extracts memory triads (e.g., Action -> Consequence) to find patterns.
- Hypothesis Generation: It compresses experience into an 8-step target plan toward external coordinates, loading it into the $\mathcal{H}_{\text{latent}}$ channel for the subcortical arbitrator to execute the next day.
Natural language is emitted solely as a unidirectional phonological output (for social dialogue and logging). It is never re-injected as raw text to drive motor arrays, eliminating syntax hallucinations.
4. Finite Somatic Memory and CLS Triage
To force continuous affective prioritization, Ascendant restricts autobiographical memory to a chronological tape of exactly $N_{\max} = 100$ engrams. Each engram is a formal 9-tuple preserving temporal, spatial, and somatic context.
Retention score $R_i(t)$ combines emotional salience, Hebbian reviviscence ($\rho_i$), and a linear temporal decay penalty:
\[R_i(t) = \left( \sigma_i + 12.0 \cdot \log(1 + \rho_i) + 4.0 \cdot \kappa_i \right) \cdot \left(1 + \vert{}\mathcal{V}_i\vert{}\right) - 0.05 \cdot (t - \tau_i)\]
Inspired by the Complementary Learning Systems (CLS) theory of the mammalian hippocampus and neocortex, the tape is partitioned into a Hippocampal Buffer (recent events) and a Cortical Zone (consolidated schemas). When the memory tape overflows, eviction occurs strictly in the cortical zone, purging the lowest $R_i(t)$ engrams. The system inherently forgets mundane walks but permanently burns in the trauma of predator attacks or the death of a parent.
5. Psychological Dynamics: Senescence and Functional Grief
Progenitors possess a telomeric limit of natural senescence ($\tau_{\text{father}} \approx 800$ cycles). Because the LLM is blind to system variables, it has no advance warning of this limit.
When a progenitor naturally dies in the presence of an offspring, the system exhibits a Simulated Grief Behavioral Response:
- An indelible high-retention trauma engram is stamped in memory.
- The ODE for Sadness ($h_{\text{tris}}$) spikes abruptly ($+0.56$), while Dopamine collapses.
- The offspring reprioritizes its goals, reducing spatial exploration and actively seeking alternative affiliative bonds (adopting orphan kin) to recover allostasis via Oxytocin.
We do not claim the Python script "feels sad." It is a functional, mathematically bound affective state associated with loss.
6. Empirical Validation: The 2x2 Factorial Lineage Benchmark
To prove the architecture's causal efficacy, we replaced projected estimations with a massive empirical benchmark executed on Engine v2.4 (Python/Node).
We tested an orthogonal 2x2 factorial protocol across four-generation lineages ($P_0 \to \text{Alfa} \to \text{Beta} \to \text{Gama}$), with 4 foraging grounds located 35 meters away, separated by rock obstacles.
- Scale: $1,440$ runs across $360$ matched seed blocks; $5,760$ individual agent biographies; $5,832,000$ cumulative simulation cycles.
- Factors: Cultural Transmission ($T^+/T^0$) $\times$ Physical Communal Pantry Storage ($P^+/P^-$).
6.1. Analysis of Variance (Two-Way ANOVA)
A Randomized Complete Block Design (RCBD) Mixed Model ($df_{\text{error}} = 1077$) isolated the seed variance and delivered definitive statistical results:
- Factor A (Cultural Transmission): Independent and highly significant cognitive contribution ($F = 29.38, p = 5.95 \times 10^{-8}, \eta_p^2 = 0.0200$). Culture alone added $+303.7$ cycles to lineage longevity, guiding trajectories safely around predators and toward known resources.
- Factor B (Communal Pantry): Massive main effect as a physical buffer against early juvenile starvation ($F = 198.58, p < 10^{-15}, \eta_p^2 = 0.1215$; $+781.6$ cycles).
- Interaction ($T \times P$): No statistical interaction was found ($p = 0.9681$). The two mechanisms operate completely independently and stack additively.
6.2. Offspring Survival Rates
Combined natural senescence survival of the offspring increased monotonically across the arms:
- 29.4% in Double Control (No culture, no pantry).
- 40.4% with Culture only.
- 54.7% with Pantry only.
- 62.6% in Full Synergy (Culture + Pantry).
6.3. Epistemic Primacy vs. Caloric Needs
In a separate single-generation study (50 seeds), we demonstrated the principle of Epistemic Primacy. When an agent ventures out, tests a hypothesis, and finds food, the "Eureka" validation engram generates a somatic retention peak of $R = 205.29 \pm 8.74$. This is statistically far superior ($t = 68.99, p < 10^{-45}$) to the retention score of basic caloric ingestion ($R = 120.00$). The system values the reduction of uncertainty more than basic satiety, driving autonomous exploration.
6.4. The Cultural Transmission Index ($K_g$)
To measure the ratcheting accumulation of knowledge across generations, we formalized the Cultural Fitness Index with a Laplace Regularizer ($+1$):
\[K_g = \frac{\sum \mathcal{V}(e)}{\vert{}\mathcal{M}_g^{\text{cult}}\vert{} + 1}\]
The $+1$ in the denominator is vital engineering: it prevents $0/0$ division errors in the ancestral generation ($P_0$) where inherited engrams are zero, and it acts as a Bayesian shrinkage prior to penalize stochastic noise in early descendants. By Generation 3 (Gama), lineages maintained a robust inherited cultural index of $1.206 \pm 0.160$.
7. Assessment Against Artificial Consciousness Frameworks
How does this map to the cutting edge of cognitive neuroscience? We evaluated Ascendant using the theoretical indicators of artificial consciousness proposed by Butlin et al. (2023, 2026):
- Global Neuronal Workspace (GNW): When a predator crosses the sensory threshold (Markov Blanket), the system broadcasts synchronized cross-module updates: Cortisol ODE spikes, a negative engram is recorded, the latent plan is purged, and the motor loop defaults to FLEE.
- Recurrent Processing Theory (RPT): Achieved via the offline daydreaming state, where autobiographical engrams are re-simulated (offline replay) to recursively alter the weights of the waking motor arbitrator.
- Higher-Order Theories (HOT): The system implements a Tripartite Epistemic Tracker. It differentiates between direct perception (certainty $\ge 0.95$), memory belief, and prospective hypotheses (certainty $< 0.40$), formally separating verified reality from internal conjecture.
- Active Inference (Predictive Processing): The agent generates a counterfactual plan, executes motor trajectories, and checks prediction error at the destination ($\varepsilon = 0$ if resource present; $\varepsilon = 1$ if absent/ambush), updating heuristic valences accordingly.
Limitation: Ascendant currently lacks a continuous Attention Schema (AST) self-model and dense sensory modalities (like retinotopic vision), operating on a discrete metric grid.
8. Phase III Roadmap: Restricted Semantic Scaffolding
A valid criticism of using Llama-3 is its reliance on pre-trained human semantic representations. The agent knows the concept of "danger" from reading human books, not from embodied experience.
In Phase III, Ascendant implements Restricted Semantic Scaffolding. We pre-train a compact foundation model exclusively on physical causality, topology, and homeostasis. This provides an initial, observable linguistic scaffold. From there, continuous LoRA adapters will update the synaptic weights based strictly on the agent's verified sensory experiences and physical traumas in the simulation, progressively overriding human semantic priors with genuine, experientially grounded representations.
9. Conclusion and Open Data Availability
Project Ascendant demonstrates that the trajectory toward synthetic agency requires thermodynamic vulnerability, strict epistemic closure, and homeostatic allostatic arbitration before LLMs can transition from statistical sequence predictors to embodied cognitive engines.
To ensure complete reproducibility and adhere to FAIR Open Science principles, the entire 554 MB SQLite relational database (ascendant_generacional_integro.db) containing the cycle-by-cycle telemetry of the $5,760$ biographies, along with the execution scripts and tabular datasets, are freely available.
We invite the software engineering and AI research community to build upon this cybernetic architecture.
Author & System Architect: David San Nicolás Montero
Full Open Source Code & Raw Data: GitHub Repository
Definitive Manuscript (Zenodo): 10.5281/zenodo.23802198
License: Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). You are free to share and adapt this material for non-commercial research, provided appropriate credit is given to the original author and DOI.