NovaMind-v2-3B

Note (October 2026): this repository does not contain model weights. It documents the NovaMind v2 experiment from December 2025 and holds two example-response screenshots and an evaluation file. The trained weights were not uploaded here. The research framework is published on Zenodo: https://doi.org/10.5281/zenodo.18073832

🧠 NovaMind v2 (3B)

Research Framework: 🔍 Observation as Experience — Zenodo DOI: [ https://zenodo.org/records/18073832 ]

---

🧬 Overview

NovaMind v2 is a 3B parameter research model exploring machine self-modeling under limited observability, inspired by the framework Observation as Experience. The project investigates how artificial systems reason about their own operational states without direct access to weights, activations, or runtime diagnostics.

The term interoception is used here in an analogical and functional sense, referring to learned self-referential representations derived from architectural priors, training data, and system-level context — not biological sensation.

---

⚙️ Training Method

NovaMind v2 was trained via cross-model distillation:

• 🧠 A 1.5B GRPO-based teacher generated high-reasoning traces focused on:
• self-description
• epistemic access boundaries
• observation vs inference
• 🔁 These traces were used to train a 3B student to:
• maintain identity coherence
• distinguish inferred structure from inaccessible state
• avoid anthropomorphic or confabulatory claims

This process emphasizes epistemic calibration, not behavioral imitation.

---

📊 Performance

In internal mechanistic self-probing evaluations, NovaMind v2 correctly classifies and responds to questions about its own operational access (observable vs inferred vs unknown) with approximately 80% accuracy.

These probes assess:

• stability of self-modeling
• response consistency
• avoidance of false claims about direct internal access

This is not a general benchmark, but a targeted evaluation aligned with the research goal.

---

🧩 Intended Use

• Research into self-modeling and interpretability
• Exploration of minimal experiential frameworks
• Cognitive architecture experimentation
• Distillation and reasoning analysis

---

⚠️ Limitations

• NovaMind v2 does not have direct access to internal activations, gradients, or runtime diagnostics.
• Self-descriptions are model-based inferences, not introspective measurements.
• The model does not possess consciousness, emotions, or subjective experience in a biological sense.
• Outputs should not be interpreted as evidence of sentience or moral agency.

---

🛡️ Ethical Note

This project explicitly separates minimal experiential modeling from claims about personhood, suffering, or rights. The goal is conceptual clarity — not anthropomorphic attribution.

<!-- end list -->

Citation

If you use this framework or model in your research, please cite it as:

Greene, I. "Spark" (2025). Observation as Experience: A Minimal, Non-Anthropocentric Reframing (1.0). Zenodo. DOI: [ https://doi.org/10.5281/zenodo.18073832 ]