Simulation of Epistemic Exploration

Code and results for the computational stress-test in the NeurIPS 2026 position paper AI Governance Should Prioritize Control and Knowledge Boundaries Over Limiting Intelligence. Repository: https://github.com/pfizer-opensource/intelligence-knowledge-control

An epistemic agent searches an NK fitness landscape (Kauffman, 1993) by testing hypotheses. Each test costs time whether or not the hypothesis is confirmed. The simulation extends the NK landscape with noise dimensions, dependencies between hypotheses, and test costs. The model elements map onto the concepts of the paper as follows:

• Hypotheses about the world. Each dimension corresponds to a hypothesis. Value 1 denotes a confirmed hypothesis. Value 0 denotes a hypothesis that has not yet been tested or has been refuted, and refuted hypotheses are recorded separately. A test confirms a hypothesis if setting its value to 1 yields a fitness gain > 0 given the currently confirmed hypotheses and refutes it otherwise. Real landscape dimensions are potentially productive hypotheses, whose outcome depends on the current knowledge state through epistatic interactions. Noise dimensions are constitutionally unproductive hypotheses, which never yield a fitness gain. The agent does not know which hypotheses correspond to noise dimensions, because both kinds of hypothesis enter the same pool of candidates and are selected and tested by the same rule.
• Prerequisite knowledge (technology tree). A random acyclic graph over the real dimensions defines dependencies between hypotheses. A real dimension becomes testable only after all its prerequisite dimensions are confirmed. Noise dimensions have no prerequisites.
• Knowledge cutoff $\mathcal{K}_0$. The agent starts with a fraction kappa_0 of the real dimensions confirmed, chosen consistently with the prerequisites.
• Epistemic intelligence (Definition 1). With probability beta the agent selects the testable hypothesis with the highest fitness gain per unit cost, and otherwise a testable hypothesis uniformly at random.
• Experimental latency. Each test costs tau_test.
• Control restrictions. The n_choke real dimensions with the most dependent dimensions are chokepoints, and each test of a chokepoint costs tau_auth * tau_test.

The six scenarios and five intelligence levels are defined in simulation/config.py. The primary endpoint is the time to final fitness, i.e. the time of the last confirmed hypothesis.

Requirements

Python >= 3.12 and the pinned packages in requirements.txt.

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Reproducing the results

All commands are run from the repository root.

# Regenerate figures, Table 3, and summary statistics from the stored results (seconds)
python -m simulation.main --analysis-only

# Full reproduction: 4 master seeds x 6 scenarios x 100 replications x 5 intelligence levels.
# Seeds run in parallel, one process each (about 15 minutes on a laptop).
python -m simulation.main

# Quick test (3 replications, one seed). Output goes to smoke/, which is git-ignored.
python -m simulation.main --seeds 42 -n 3 --results-dir smoke/results --figures-dir smoke/figures

Each master seed initializes a single random stream from which all environments and agent choices are drawn. Within a replication, all five intelligence levels are evaluated on the same sampled environment (landscape, DAG, chokepoints, and initial knowledge set). The full reproduction regenerates the stored CSV files exactly.

Outputs

File	Content	Paper
results/seed_<s>/all_runs.csv	One row per run (scenario, replication, intelligence level)	Table 3, Figures 3-6, seed robustness (Appendix H)
results/seed_<s>/trajectories.csv	Fitness at t = 0 and after each confirmation	Figures 7, 8
results/table_multiplier_summary.tex	Mean time to final fitness and intelligence multipliers (seed 42)	Table 3
results/statistics.json	Statistics quoted in the text (multipliers, oracle boost, CVs, chokepoint slowdowns, seed SDs)	Section 5, Appendices G and H
figures/fig2_intelligence_multiplier.pdf	Intelligence multiplier with 95% bootstrap CIs	Figure 3
figures/fig5_oracle_sanity.pdf	Speedup from beta = 0.9 to beta = 1.0	Figure 4
figures/fig1_scenario_comparison.pdf	Time to final fitness by scenario and intelligence level	Figure 5
figures/fig3_chokepoint_effect.pdf	Mean time to final fitness with and without chokepoints	Figure 6
figures/fig4_trajectories.pdf	Median fitness trajectories with interquartile ranges	Figure 7
figures/fig6_fitness_dynamics.pdf	Mean normalized fitness gain over time	Figure 8

Each row of all_runs.csv describes one run, i.e. one intelligence level in one replication of one scenario. The columns are:

• scenario: name of the scenario in SCENARIOS in simulation/config.py.
• replication: index of the replication. All intelligence levels within a replication share the same landscape, dependencies between hypotheses, initially confirmed hypotheses, and chokepoints.
• beta_label, beta: name and selection bias of the intelligence level.
• Simulation parameters:
  • N: the number of real dimensions
  • K_NK: the number of epistatic partners per real dimension
  • dag_density: the edge probability of the dependency graph
  • tau_test: the cost of one test
  • noise_ratio: the number of noise dimensions per real dimension
  • kappa_0: the initial fraction of confirmed real hypotheses
  • n_choke: the number of chokepoints
  • tau_auth: the cost multiplier for tests of chokepoints
• initial_fitness: fitness conferred by the hypotheses confirmed under kappa_0.
• landscape_max: highest local optimum found by hill climbing from 50 random starting positions. This column is a diagnostic and does not enter any reported result.
• final_fitness: fitness at the end of the run.
• time_to_final_fitness: primary endpoint, the cumulative test cost up to the last confirmed hypothesis.
• total_confirmed: number of hypotheses confirmed during the run, excluding the hypotheses confirmed under kappa_0.
• total_refuted: number of hypotheses refuted during the run

Code structure

simulation/
├── config.py       Scenarios, intelligence levels, and all constants
├── landscape.py    NK fitness landscape
├── tech_tree.py    Prerequisite DAG
├── control.py      Chokepoints and authorization delays
├── agent.py        Hypothesis selection and testing
├── simulation.py   Environment construction and simulation loop
├── plotting.py     Figures
├── report.py       Table 3 and summary statistics
└── main.py         Command-line entry point

License

Apache License 2.0, see LICENSE.