Social Sandbox

Thin scenario wrapper for counterfactual social simulation. Differs from product-feedback (scores products) and vote-predict (categorical votes) — this produces open-ended qualitative narratives from the panel, then clusters them.

Recipe

import sys
sys.path.insert(0, str(__import__('pathlib').Path.home() / '.claude/skills/persona-sim'))
from lib import sampler
from lib.sim_engine import SYSTEM_PROMPT, _persona_card
from lib.llm_router import generate
from concurrent.futures import ThreadPoolExecutor

# 1. Sample diverse panel (wider demographic spread than product-feedback)
panel = sampler.sample_personas(n=50, source="nemotron_usa", mode="stream")

# 2. Design the scenario — use past tense as if it already happened
scenario = """
Congress has just passed a law raising the federal minimum wage to $22/hour
nationwide, effective in 6 months. You've seen the news today.
"""

# 3. Ask each persona: immediate reaction + 6-month expectation + what they plan to do
def probe(persona):
    task = (f"{scenario}\n\n"
            "Respond as yourself, in 3 sentences:\n"
            "(1) Your immediate emotional reaction.\n"
            "(2) What you expect to happen in your life over 6 months.\n"
            "(3) What (if anything) you plan to do in response.")
    return generate(system=SYSTEM_PROMPT, persona_card=_persona_card(persona),
                    task=task, tier="default", max_tokens=400)

with ThreadPoolExecutor(max_workers=8) as ex:
    narratives = list(ex.map(probe, panel))

# 4. Cluster the narratives (use Sonnet for this, not Haiku)
# Pass the 50 narratives + panel demographics to Sonnet:
# "Identify 3-5 distinct reaction archetypes. For each: label, % of panel, key demographic correlates,
#  sample quote."