[fdrocha]

Stacked on #29 (Micropolis: natural-conditional study), which is stacked on #28. The commits new in this PR are: c90b903 Micropolis: interventions in run_sim and ground-truth continuations; 548b4e1 Micropolis: interventional-conditional draw; 50689e5 Micropolis: per-continuation interventions and interventional-conditional prompts; 17097fd Micropolis intcond: say a forced disaster is reported once, when it starts; e920d8a Micropolis intcond: lint fixes.
Diff of just these commits: fdrocha/forecastbench-sim@micropolis-cond...micropolis-intcond

An interventional-conditional (intcond) study for Micropolis, P(A | do X), after Starsim's but drawn like the natural conditionals of #29: an intervention X is applied right after the snapshot, the snapshot is replayed 1,000 times with it, and models are asked question A before and after being told of X.

• Interventions in the simulator. CitySimulation takes (turn, action) interventions (run_sim.py --at); ids and file names gain an -at<T>_<action> suffix only when present, so baseline cache keys are unchanged. extract_ground_truth.py --intervene writes <scenario>_T<S>_do-<action>.jsonl. The intervention is applied per continuation, after its reseed (MicropolisCore fbs 4ba571d), so each replay draws its own spot for the disaster, as the prompt says; a miss is retried with another seed and recorded (missed_seeds), and 100 misses mark the action infeasible (--skip-infeasible). A failed engine run now reports its stderr.
• Seven actions: earthquake, fire, flood, tornado, meltdown, demolish all power plants, demolish the airport. No partial demolitions; the monster is left out (the engine spawns it on water, where it disappears).
• Draw. draw_intcond.py draws 400 cells from the base and do-X ground truth at S = 960 and 1440, in four structural blocks: null 50 (disaster → unrelated question), same-family 50 (disaster → later disasters of the same kind), damage 200, infrastructure 100. 0.05 < P(A) ≤ 0.95 and 0.02 < P(A|do X) < 0.98; weak action–template pairs are dropped by family from a pilot, never per cell. 34 of 210 city/snapshot/action combinations are infeasible (no airport 12, no nuclear plant 14, flood 8).
• Prompts and elicitation. build_intcond_prompts.py: turn 1 is the natural-conditional turn 1 verbatim, so elicit_natcond.py --study intcond copies those replies (copied_from) instead of asking again; turn 2 (item <question_id>|do-<action>) describes the intervention and restates A. For a forced disaster, the note says the disaster "is reported once, when it starts … and is not reported again while it lasts", after Llama and Gemini counted the forced disaster itself under the first wording.

Testing

• pytest worlds/micropolis/tests: 266 passed (new test_intcond.py; test_ground_truth.py extended for interventions and infeasible markers). Ruff is clean on the new files.
• Ground truth: do-files for all feasible city/snapshot/action combinations, 1,000 continuations each; miss rates ≤ 20% for flood, ≤ 2% for fire, 0 for the others.
• Full elicitation: 22 models × 400 cells, both turns ($104.70 of new calls). Turn-2 replies whose messages changed with the rewording were re-asked automatically (Llama 4 Scout: 286), as was a GPT-5 Mini turn 2 built on a cut-off turn 1. All 15,950 scored answers match the current prompts' message hashes. Results are in the paper repo (micropolis_intcond.py) and the data archive (datasets/micropolis-intcond/).
• Not tested: CI does not install the engine, so the engine-side intervention code is tested only locally.

🤖 Generated with Claude Code

https://claude.ai/code/session_015Vag3y8U9eNWU2f51EPoPW