Behavioral-Coherence Interventions

The examples/exp_behavioral_coherence experiment provides six intervention conditions derived from one common baseline. The *_llm.json variants keep the baseline population and environment settings and use openai/gpt-oss-120b for the single LLM household.

Conditions

Condition	Configuration	Intervention
E1	config_e1_price_increase_llm.json	DiscountRetailer raises Daily Mart staple prices by 10 percent from step 144. It uses the same time lookup and discounts schema as the discount-restaurant example; a negative discount is a markup.
E2	config_e2_supply_reduction_llm.json	Daily Mart is removed from ConstantSupply and moved to DynamicSupply. Its replenishment ratio changes from 0.5 to 0.25 at step 144.
M1	config_m1_keep_out_llm.json	KeepOut prevents entry into the four common-space cells from step 144 through the end of the run.
M2	config_m2_fast_car_llm.json	A same-price HighPerformanceCar with velocity 4 becomes available at step 144; the baseline GasolineCar has velocity 3.
S1	config_s1_positive_info_llm.json	Rule-based households may replace an ordinary tweet with positive restaurant information from tweet_candidates.
S2	config_s2_misinformation_llm.json	Rule-based households may replace an ordinary tweet with false Daily Mart stockout information from tweet_candidates.

DiscountRetailer is experiment-specific and lives in intervention_agents.py. DynamicSupply and KeepOut are reusable built-in events documented in Configuration Reference.

Running a condition

Run commands from the repository root. For example, run E1 once with seed 42:

.venv/bin/python examples/exp_behavioral_coherence/main.py \
    --config examples/exp_behavioral_coherence/config_e1_price_increase_llm.json \
    --condition e1_price_increase \
    --model-label gpt-oss-120b \
    --seed-start 42 \
    --num-seeds 1

--config determines simulation behavior. --condition and --model-label only determine the output directory; omitting --condition uses baseline and can overwrite a baseline log with the same seed. Results are written to:

examples/exp_behavioral_coherence/logs/<condition>/<model-label>/<seed>.txt

The no-argument default deliberately runs the non-LLM config_baseline.json so a routine local check cannot start a vLLM server accidentally. Add --summarize to print the simulation configuration summary. The LLM files configure GPU IDs 0 and 1 and allow VLLMClient to start its server as usual.

Candidate paths

Candidate-file paths in JSON are relative to the configuration file. The experiment runner resolves them to absolute paths before creating the simulator, so invocation from another working directory does not change which file is loaded.