Procela: Epistemic Governance in Mechanistic Simulations Under Structural Uncertainty

Kinson Vernet

physics.comp-phApr 1, 2026

Procela: Epistemic Governance in Mechanistic Simulations Under Structural Uncertainty

Authors: Kinson Vernet

Organizations: Independent Researcher

Abstract

Mechanistic simulations typically assume fixed ontologies: variables, causal relationships, and resolution policies are static. This assumption fails when the true causal structure is contested or unidentifiable-as in antimicrobial resistance (AMR) spread, where contact, environmental, and selection ontologies compete. We introduce Procela, a Python framework where variables act as epistemic authorities that maintain complete hypothesis memory, mechanisms encode competing ontologies as causal units, and governance observes epistemic signals and mutates system topology at runtime. This is the first framework where simulations test their own assumptions. We instantiate Procela for AMR in a hospital network with three competing families. Governance detects coverage decay, policy fragility, and runs structural probes. Results show 20.4% error reduction and 69% cumulative regret improvement over baseline. All experiments are reproducible with full auditability. Procela establishes a new paradigm: simulations that model not only the world but their own modeling process, enabling adaptation under structural uncertainty.

Figures & tables

Figure 1: Mechanism contributions (governance combined) for the AMR case study. The predicted colonized patients is for combined governances (Section 5 ).

Figure 2: Regime shifts for the AMR case study.

Metric	Value	95% CI
Mean absolute error	0.535	[0.516, 0.554]
Standard deviation	0.383	[0.370, 0.396]

Table 1: Baseline prediction error with no governance.

Figure 3: Coverage timeline for the baseline.

Metric	Value	Improvement w.r.t baseline
Mean absolute error	0.544 $\pm$ 0.019	-1.636%
Standard deviation	0.386 $\pm$ 0.013	-0.712%
Experiments triggered	7 $\pm$ 0	–
Successful experiments	5 $\pm$ 0	71.4% success rate

Table 2: Policy fragility governance results in comparison to the baseline.

Figure 4: Timeline of policy fragility experiments and confidence share between ontologies. Shaded regions indicate experiment periods for 1 experiment. Successes (green) occur in high-error regimes.

Metric	Value	Improvement w.r.t baseline
Mean absolute error	0.426 $\pm$ 0.036	20.42%
Standard deviation	0.327 $\pm$ 0.027	14.76%
Experiments triggered	11 $\pm$ 0	–
Successful experiments	6 $\pm$ 0	54.5% success rate

Table 3: Coverage decay governance results.

Figure 5: Experiment outcomes for coverage governance. (Left): coverage for baseline (no governance), (Right): coverage for coverage decay governance. By comparison to the results with no governance, coverage shows that contact family is continuously disabled from the selection regime.

Figure 6: Absolute error and confidence share over time for the three families under coverage decay governance. Shaded regions indicate when a family was disabled. Coverage shows a total of six successful experiments over 11 triggered during different regimes.

Metric	Value	Improvement w.r.t baseline
Mean absolute error	0.485 $\pm$ 0.020	9.31%
Standard deviation	0.364 $\pm$ 0.018	5.00%
Probes triggered	3 $\pm$ 0	–
Successful probes	2 $\pm$ 0	66.67%

Table 4: Structural probe governance results in comparison to the baseline.

Figure 7: Isolated performance for each family across probes. Isolation is chosen to avoid total coincidence to regime. During selection regime, contact family is isolated to estimate its performance. During environmental regime, environmental family is isolated. During contact regime, selection family is isolated.

Figure 8: Absolute error and confidence share for structural probe over time. Shaded regions indicate when a family was isolated. Probe shows two successful experiments during environmental and contact regimes (family errors are higher) and 1 failed experiment duration selection regime (family errors are lower).

Metric	Value	Improvement w.r.t baseline
Mean absolute error	0.491 $\pm$ 0.030	8.18%
Standard deviation	0.378 $\pm$ 0.034	1.31%

Table 5: Combined governances results in comparison to the baseline.

Figure 9: Performance for the combined governances. (Top, left): intervention code over time shows how combined governances tries to reconstruct the intervention despite including an ineffective governance component; (Top, right): absolute prediction error over time; (Bottom, left): coverage indicates that a family might be not completely isolated; (Bottom, right): policy fragility over time.

Figure 10: Performance timeline of the combined governances. Almost all experiments failed during the selection regime where family coverages are high. When family coverages are low (environmental and contact), experiments are likely to be succeeded.

Strategy	Mean error	Cumulative difference
Policy fragility	-1.63%	-9.8%
Coverage decay	20.42%	35.5%
Structural probe	9.31%	69.0%
Combined governance	8.18%	61.6%

Table 6: Decision-making performance through cumulative difference in comparison to the baseline.

Figure 11: (Left): Comparative performance of governance strategies for our AMR case study. (Right): Cumulative difference relative to the optimal intervention (intervention code = 0) over time. Coverage decay makes better predictions (+20.42%) while structural probe makes better decisions (C.E=+69.0%).

Figure 12: Topology evolution for each governance in the AMR study. The contact family has three mechanisms, while the environmental and selection families have four each. In total, the AMR case study has eleven competing mechanisms.

Strategy	Mean error	Std error	Error	C.R
No governance	0.535 $\pm$ 0.019	0.383 $\pm$ 0.013	–	–
Policy fragility	0.544 $\pm$ 0.019	0.386 $\pm$ 0.013	-1.63%	-9.8%
Coverage decay	0.426 $\pm$ 0.036	0.327 $\pm$ 0.027	20.42%	35.5%
Structural probe	0.485 $\pm$ 0.020	0.364 $\pm$ 0.018	9.31%	69.0%
Combined	0.491 $\pm$ 0.030	0.378 $\pm$ 0.034	8.18%	61.6%

Table 7: Comparative performance of governance strategies.

Explore similar work

Simulate, Reason, Decide: Scientific Reasoning with LLMs for Simulation-Driven Decision Making
Jun 3, 2026Yuhan Yang, Ruipu Li, Alexander RodríguezAI for Science Neuro-Symbolic Reasoning
Computer Science and Engineering University of Michigan
Science Utopia? Closed-Loop LLM Simulation of Academic Research Ecosystems
Oct 1, 2026Yiqiao Jin, Yiyang Wang, Lucheng Fu +8Scholarly Peer Review Computational Social Science
Georgia Institute of Technology · University of California, Los Angeles · University of Chicago +1
Stop Drawing Scientific Claims from LLM Social Simulations Without Robustness Audits
May 17, 2026Jinyi Ye, Lei Cao, Ding Chen +1Human Behavior Simulation Agent-Based Modeling
Thomas Lord Department of Computer Science, University of Southern California · Annenberg School for Communication and Journalism, University of Southern California · Marshall School of Business, University of Southern California

Jun 3, 2026J/K move · Enter open · S save

Simulate, Reason, Decide: Scientific Reasoning with LLMs for Simulation-Driven Decision Making

Yuhan Yang, Ruipu Li, Alexander Rodríguez

Computer Science and Engineering University of Michigan

AI for Science Neuro-Symbolic Reasoning LLM Reasoning

Open paper AskPDF arXiv

Scientific simulators are increasingly being integrated into LLM-driven systems for high-stakes simulation-driven decision-making. However, existing frameworks primarily use LLMs to generate, calibrate, or execute simulators, treating them as black-box interfaces rather than as structured mechanistic systems that can be reasoned about. As a result, current approaches lack the ability to identify, represent, and reason about the assumptions and mechanisms underlying simulator behavior, limiting transparency, auditability, and decision justification. We introduce MechSim, a mechanism-grounded neuro-symbolic reasoning framework for executable scientific simulators. Unlike prior neuro-symbolic approaches that primarily reason over static symbolic structures, MechSim enables LLM agents to reason about the mechanisms, assumptions, and execution behavior of scientific simulators. Our framework represents simulators through a shared structured schema capturing assumptions, variables, mechanism dependencies, and execution traces. On top of this representation, LLM agents operate as constrained reasoning engines that generate structured, evidence-grounded explanations linking simulator outcomes to their underlying mechanisms. We evaluate our approach across multiple high-stakes domains and show that it improves mechanism-level explanation quality, simulator analysis, and downstream decision-making reliability.

Abstract

Explore similar work

Simulate, Reason, Decide: Scientific Reasoning with LLMs for Simulation-Driven Decision Making

Science Utopia? Closed-Loop LLM Simulation of Academic Research Ecosystems

Stop Drawing Scientific Claims from LLM Social Simulations Without Robustness Audits