Communication Simulation
Overview
This Skill treats simulation as an instrument for testing whether specified micro-rules
can reproduce macro communication patterns. It does not treat plausible model output as
evidence that the rules are true.
Required start
- State the macro phenomenon, proposed micro mechanisms, agents, environment,
interaction topology, schedule, and target pattern.
- Write the simulation contract in
references/simulation-protocol.md.
- Separate explanation, prediction, measurement, and scenario exploration.
Method routing
- classical ABM encodes explicit state-transition and interaction rules;
- opinion-dynamics models require justified update, confidence, and network rules;
- generative or LLM agents require model/version, prompt, memory, tool, sampling, and
refusal/error provenance;
- microsimulation reproduces population composition but not interaction mechanisms by
itself;
- system dynamics models aggregate stocks and flows and cannot answer actor-level
questions without a bridge.
Validation
Calibrate independent parameters to empirical data, reserve separate targets for
validation, compare against simple/null models, run repeated seeds, report stochastic
uncertainty, sweep plausible parameter ranges, test initialization/topology/schedule,
and examine equifinality. For LLM agents, compare absolute levels, rank ordering,
mechanisms, heterogeneity, and group structure against human reference data.
Output shape
Return the mechanism map, simulation contract, calibration/validation split, parameter
table, baseline models, experiment grid, seed and uncertainty plan, sensitivity and
equifinality results, and a bounded statement of what the simulation explains.
Hard boundaries
- Macro pattern reproduction does not identify a unique micro mechanism.
- Prompt tuning is not human calibration.
- Agent personas are not sampled populations.
- One seed or one attractive run is not a result.
- Model agreement with one target can coexist with failure on mechanisms and subgroups.
1---2name: communication-simulation3description: Design and audit agent-based, generative, and LLM-agent simulations of communication processes with explicit mechanisms, calibration, sensitivity, and validation. Use for ABM/生成式仿真/LLM智能体/合成公众/意见动力学. Do not present simulated agents as human evidence without empirical calibration.4---56# Communication Simulation78## Overview910This Skill treats simulation as an instrument for testing whether specified micro-rules11can reproduce macro communication patterns. It does not treat plausible model output as12evidence that the rules are true.1314## Required start15161. State the macro phenomenon, proposed micro mechanisms, agents, environment,17 interaction topology, schedule, and target pattern.182. Write the simulation contract in19 [references/simulation-protocol.md](references/simulation-protocol.md).203. Separate explanation, prediction, measurement, and scenario exploration.2122## Method routing2324- classical ABM encodes explicit state-transition and interaction rules;25- opinion-dynamics models require justified update, confidence, and network rules;26- generative or LLM agents require model/version, prompt, memory, tool, sampling, and27 refusal/error provenance;28- microsimulation reproduces population composition but not interaction mechanisms by29 itself;30- system dynamics models aggregate stocks and flows and cannot answer actor-level31 questions without a bridge.3233## Validation3435Calibrate independent parameters to empirical data, reserve separate targets for36validation, compare against simple/null models, run repeated seeds, report stochastic37uncertainty, sweep plausible parameter ranges, test initialization/topology/schedule,38and examine equifinality. For LLM agents, compare absolute levels, rank ordering,39mechanisms, heterogeneity, and group structure against human reference data.4041## Output shape4243Return the mechanism map, simulation contract, calibration/validation split, parameter44table, baseline models, experiment grid, seed and uncertainty plan, sensitivity and45equifinality results, and a bounded statement of what the simulation explains.4647## Hard boundaries4849- Macro pattern reproduction does not identify a unique micro mechanism.50- Prompt tuning is not human calibration.51- Agent personas are not sampled populations.52- One seed or one attractive run is not a result.53- Model agreement with one target can coexist with failure on mechanisms and subgroups.
haochengw372-hash