Enactive Inference Model

This repository implements a three-layer active-inference model of focused-attention meditation. It simulates expert and novice phenotypes across the canonical cycle of breath focus, mind wandering, meta-awareness, and redirecting attention, then produces publication figures and saved JSON results for analysis.

The model couples:

• Layer 1 neural-network dynamics over DMN, VAN, DAN, and FPN
• Layer 2 latent thoughtseed inference and policy evaluation
• Layer 3 metacognitive monitoring, habit priors, and meta-awareness

One run per phenotype consists of 12,000 contiguous timesteps:

• Train: 8,000 steps with learning enabled
• Eval: 2,000 frozen steps
• Plot: 2,000 frozen steps used for Fig. 3-Fig. 5 analyses

Fig. S1 is generated from the full run with eval and plot windows shaded.

---

Quick Start

Run commands from the repository root (where run_enactive_inference.py lives). If you use uv run python run_enactive_inference.py from another directory, Python looks for that path relative to your current working directory and will fail.

Create the project environment and install dependencies (including PyTorch):

uv sync

Run the full pipeline (train, save JSON, generate figures). The subcommand defaults to run when omitted:

uv run python run_enactive_inference.py

Equivalent explicit forms:

uv run python run_enactive_inference.py run
uv run python -m run_enactive_inference
uv run python -m run_enactive_inference run

Generate plots only from previously saved results under data/ (requires training_results_expert_seed42.json and training_results_novice_seed42.json from a prior full run):

uv run python run_enactive_inference.py plot

Troubleshooting: If you see ModuleNotFoundError: No module named 'torch', you are not using the project environment. Use uv sync then uv run ... as above, or activate .venv after uv sync and run the same commands with python instead of uv run python.

This project runs on CPU only; no GPU support is required or used.

---

Architecture

• Layer 1 is a neural generative process over four large-scale networks (DMN, VAN, DAN, FPN) with four meditation regimes (BF, MW, MA, RA) and multivariate Ornstein-Uhlenbeck dynamics.
• Layer 2 compresses network dynamics into five thoughtseeds, performs encoder/decoder inference, and evaluates expected free energy over stay/switch policy candidates.
• Layer 3 computes meta-awareness from policy–habit conflict (KL-based), gated by thoughtseed “ignition,” then selects the policy posterior used for descending predictions.
• L1 and L2 interact through a Markov blanket carrying network activity and dwell progress upward, and descending network predictions plus policy-state probabilities downward.
• L2 and L3 interact through a Markov blanket carrying state belief, policy evidence, and thoughtseed activations upward, and sensory precision downward.

Model Summary

• States: BF, MW, MA, RA
• Networks: DMN, VAN, DAN, FPN
• Thoughtseeds: attend_breath, pain_discomfort, pending_tasks, aha_moment, equanimity

---

Outputs

Saved Results

Running the full pipeline writes one JSON file per phenotype to data/:

• training_results_expert_seed42.json
• training_results_novice_seed42.json

Each file contains the full 12,000-step run:

• state history
• network activation history
• thoughtseed activation history
• thoughtseed prior activation history
• free-energy history
• loss history
• meta-awareness history
• raw transition events

Generated Figures

Plots are written to figures/.

Fig. S1: Convergence diagnostics

• FigS1_Convergence_Expert.pdf
• FigS1_Convergence_Novice.pdf
• Full-run diagnostics with eval and plot windows shaded

Fig. 3: Expert vs novice comparison

• fig3a.pdf: state-conditional network activation profiles
• fig3b.pdf: dwell-time comparisons by state
• fig3c.pdf: transition probability matrices from the plot window

Fig. 4: Hierarchical dynamics

• fig4a.pdf: novice hierarchical traces
• fig4b.pdf: expert hierarchical traces
• Shows L3 meta-awareness, L2 thoughtseed trajectories, and L1 network dynamics

Fig. 5: State-space geometry

• fig5.pdf: pooled PCA projections of L2 thoughtseed and L1 network trajectories

Plot window convention

• Fig. 3-Fig. 5 and dwell/transition summaries use the final 2,000-step plot window.
• Fig. S1 spans the full run with eval (steps 8,000-10,000) and plot (steps 10,000-12,000) windows highlighted.

---

Repository Layout

.
+-- run_enactive_inference.py      # Main entry point (run | plot)
+-- model/
|   +-- training_loop.py           # Variational EM loop and result packaging
|   +-- phenotype.py               # Expert/novice phenotype definitions
|   +-- l1_generative_process.py   # Layer 1 MVOU dynamics
|   +-- l2_recognition.py          # Layer 2 inference, decoder, forward model, EFE
|   +-- l3_metacognition.py        # Layer 3 meta-awareness and policy selection
|   +-- markov_blankets.py         # Markov blanket interfaces
+-- utils/
|   +-- config.py                  # Core constants and phenotype tables
|   +-- math_utils.py              # Shared tensor/math helpers
|   +-- extract_stats.py           # CLI stats summary from saved JSON
+-- viz/
|   +-- analysis_utils.py          # Tail-window statistics and aggregation
|   +-- convergence.py             # Fig. S1
|   +-- radar_plot.py              # Fig. 3A
|   +-- hierarchy.py               # Fig. 4
|   +-- attractors.py              # Fig. 5
|   +-- plotting_utils.py          # Shared plotting style/helpers
+-- data/                          # Saved run outputs
+-- figures/                       # Generated and manuscript figure assets
+-- tests/
    +-- test_invariants.py         # Core invariants and numerical checks

---

Configuration

Edit utils/config.py to modify:

• network/state parameters (THETA_BASE, network attractors)
• thoughtseed priors (THOUGHTSEED_STATE_PRIORS)
• dwell ranges (DWELL_TIMES)
• transition priors (STATE_TRANSITION_PROBS)
• learning rates (LEARNING_RATES)
• numerical settings such as NOISE_LEVEL, BPTT_STEPS, TRAIN_STEPS, EVAL_STEPS, and PLOT_STEPS

---

Reproducibility

The default seed is fixed at 42.

• torch.manual_seed and np.random.seed are set in the trainer
• Layer 1 uses its own seeded RandomState
• stochastic transitions and process noise are therefore reproducible under the same configuration and dependency environment

---

Citation

If you use this model in your research:

@article{enactive_inference_thoughtseeds_2026,
  author = {Kavi, P. C. and Friedman, D. A. and Patow, G.},
  title = {Thoughtseeds as Latent Causes: A Computational Phenomenology of Focused-Attention Meditation},
  journal = {Proc. R. Soc. A},
  year = {2026}
}

This repository builds on earlier Thoughtseeds-related codebases, including:

• https://github.com/prakash-kavi/thoughtseeds_vipassana
• https://github.com/prakash-kavi/viapssana_ts2
• https://github.com/prakash-kavi/aif_iwai2025_thoughtseeds