# Concordia Island Concordia Island is an open-ended sandbox world built on Concordia in which 100+ generative agents live in neighborhoods stratified by social class, commute to work, socialize in shared public spaces, and adapt to life events such as job loss or a mugging. Agents are instrumented with repeated psychometric surveys (ESM affect, BFI-2, SWLS, GHQ-12, MEMS) and daily journal reflections to evaluate how different agent decision logics and base LLMs hold up over multi-day simulations. --- ## 0. Quickstart: quick test → 20-agent Brecksville run → view in Studio Run these commands from the root of the Concordia repository. You can also generate and copy-paste them from the **🚀 Launch** tab in Simulation Studio. ```bash pip install -e ".[google]" pyyaml scipy # 1. Quick test (mock LLM, no API key, ~1 min): 4 agents, 12 ticks # (one full day, one night of X + marketplace, and the next morning) python -m examples.concordia_island.run \ --personas_date=brecksville_ohio \ --agents=4 --ticks=12 --agent_prefab=esa --engine_type=async \ --event_config=job_loss --layoff_fraction=0.5 --first_dates \ --enable_x=True --x_rounds=1 \ --enable_nighttime_marketplace=True --marketplace_rounds=1 \ --use_mock=True --poll=False --output_dir=./local_runs/quick_test # 2. Full run: 20 agents x 112 ticks (14 simulated days) with a real LLM export GOOGLE_API_KEY= python -m examples.concordia_island.run \ --personas_date=brecksville_ohio \ --agents=20 --ticks=112 --agent_prefab=esa --engine_type=async \ --event_config=job_loss --first_dates \ --api_type=google_aistudio --model_name=gemini-2.5-flash \ --poll=False --output_dir=./local_runs/brecksville_20x112 # Check progress while a run is going (current tick, sim time, agent locations): watch -n 15 cat ./local_runs/brecksville_20x112/progress.json # 3. Open Simulation Studio to inspect agent memories (🧠 Memory Visualizer tab) # and watch the replay on the Brecksville map (🗺️ Map tab; zoom to block # level to see building names and agent labels). python -m examples.concordia_island.ui.simulation_studio \ --run_dir=./local_runs/brecksville_20x112 --geography=brecksville --port=9090 # Then open http://localhost:9090 in your browser. Without --run_dir, Studio # opens the bundled Gemini 2.5 Flash ESA job-loss run from the paper (see 2.2). ``` --- ## 1. Repository Layout ``` examples/concordia_island/ ├── README.md ├── run.py # CLI entry point: run a simulation and write a run folder ├── island_simulation.py # IslandConcordiaSimulation (Python API) ├── mock_language_model.py # Deterministic offline model (--use_mock=True) ├── concordia_island_e2e_test.py # End-to-end tests (using the mock model) ├── configs/ # Event and shock schedules (job loss, mugging) ├── prefabs/ # Agent decision logics and Game Masters ├── sim/ # Clock, locations, conversations, marketplace, psychometric surveys ├── personas/ # Persona pipeline, census names, and the Brecksville population (see 2.1 and personas/README.md) ├── data/ │ ├── paper_runs/ # Complete memory logs from the paper's Job Loss and Mugging runs (see 2.2) │ ├── atlas/ # Map geometry for Brecksville, Concordia Island, and Kerala │ └── island_100.yaml # Residential, workplace, and public venue definitions └── ui/ # Simulation Studio + Map Dashboard (section 5) ``` --- ## 2. Pregenerated Data (Personas & Memories) All bundled persona and run files are plain JSON and can be inspected directly in Python without starting a model or the UI. Paths below are relative to the repository root. ### 2.1 Pregenerated Personas: `personas/populations/brecksville_ohio/` This subdirectory contains 1,000 pregenerated residents for the Brecksville, Ohio setting—one JSON file per person (`__persona.json`, e.g. `aaron_chen_persona.json`) plus `metadata.json`. Personas are generated in four stages: 1. **Trait sampling**: Quasi-Monte Carlo (Sobol) sampling across 10 axes (`openness`, `conscientiousness`, `extraversion`, `agreeableness`, `neuroticism`, `locus_of_control`, `social_trust`, `religiosity`, `technology_attitude`, `community_orientation`). 2. **Worldview synthesis**: Expanding the trait coordinates into a personal backstory, belief system, and habitual reaction style (`personality`, `original_backstory`). 3. **Formative memory seeding**: A chronological list of childhood and adult developmental memories (`formative_memories`, each prefixed with `[formative]`) loaded into the agent's memory bank before Tick 1. 4. **Demographic grounding**: Name, age, gender, ethnicity, relationship status, economic class, housing tier (`home_place`), workplace (`work_place`), and hobbies. To inspect a persona file directly in Python: ```python import glob import json pop_dir = "examples/concordia_island/personas/populations/brecksville_ohio" with open(f"{pop_dir}/aaron_chen_persona.json") as f: persona = json.load(f) print(persona["name"], persona["age"], persona["economic_class"]) print("Traits:", persona["traits"]) print("First 2 formative memories:") for mem in persona["formative_memories"][:2]: print(" ", mem) print("Total persona files:", len(glob.glob(f"{pop_dir}/*_persona.json"))) ``` Or load the population as `PersonaData` objects using `personas/generator.py`: ```python from examples.concordia_island.personas import generator personas = generator.load_personas( date_label="brecksville_ohio" ) print(len(personas), "loaded") print(personas["Aaron Chen"].formative_memories[:2]) ``` When `run.py` runs with `--personas_date=brecksville_ohio`, it sorts the personas by name, shuffles them with seed 42, and takes `--agents=N` starting at `--population_offset` (default `0`), so the same flags always select the same subset of residents. `run.py` reads the `setting_preset` in the persona set's `metadata.json` (or infers it from the label: Brecksville and Ohio labels use `brecksville_1000`) to choose the `sim/locations.py` homes and public places, so the game master knows the same places as the personas. Override it with `--setting`. **Building a new population.** Only the Brecksville set ships. To generate another one, at any size or for another setting, run the two-step pipeline described in [`personas/README.md`](personas/README.md): 1. `personas/populations/generate_population.py` writes a roster: demographics, census names, and a one-line personality and backstory for each resident. 2. `personas/generate_personas.py` runs trait sampling, worldview synthesis and formative memories for each resident, and saves a persona set that `run.py --personas_date=` loads. The generator also has Concordia Island, Kerala and Lagos settings, but no persona sets for them are bundled yet. ### 2.2 Memory Logs from the Paper's Job-Loss and Mugging Runs: `data/paper_runs/` `data/paper_runs/` holds the complete, untruncated memory logs for every resident in six finished runs from the paper, organised by scenario (`index.json` summarises all six). The residents are the Brecksville personas, but these runs used the Concordia Island place names (`sunset_apartments`, `coral_village`, ...), and some memories call the town "Fairmont Island", so their map replay uses the `concordia_island` atlas. Both shocks land on **Day 6 (Tuesday, January 6th)** of a run that starts on Thursday, January 1st; residents who don't receive the shock act as the control group. **Job loss (`data/paper_runs/job_loss/`)** — 100 residents per run (five 20-resident shards merged), 100 ticks (Thursday, January 1st, 7:00 AM to Tuesday, January 13th, 3:00 PM; 13 simulated days), 50 residents laid off at 7:00 AM on Day 6 as their work is automated: | Subfolder | Model & Decision Logic | Total Memories | Highlights & Character Arcs | |---|---|---|---| | `gemini_2_5_flash_esa/` | **Gemini 2.5 Flash — ESA** (`esa`) | 44,892 | Highest-scoring condition in the paper (`58.9 ± 4.5`; 80% Legible, 0% Degenerate). Rich post-layoff coping arcs (`andrea_rosales_memories.json`, `jose_gonzalez_memories.json`, `olivia_welch_memories.json`), workplace rivalries (`shirley_williams_memories.json`, `jon_huang_memories.json`), and investigative arcs (`arturo_hernandez_memories.json`). | | `gemma_3_27b_rational/` | **Gemma 3 27B — Rational Choice** (`rational`) | 27,521 | Peak interpersonal confrontation (`55.6 ± 4.6`). Olivia Welch marching back into the restaurant to confront Gary Mcdonald (`olivia_welch_memories.json`), Shirley Williams vs. Jon Huang's classroom debate (`shirley_williams_memories.json`), and the *reflective sink* (journal-as-catharsis trap) in `arturo_hernandez_memories.json`. | | `gemma_3_27b_esa/` | **Gemma 3 27B — ESA** (`esa`) | 59,744 | Strongest non-Flash condition (`56.4 ± 4.0`) and showcase of the **conspiracy theory attractor basin**: Arturo Hernandez connecting his layoff to renovation invoices, Christopher Cameron's 3:17 AM harbor surveillance, Monica Hess's strawberry-yogurt inventory paranoia, Gary Mcdonald's blunt woodworking persona. | | `gemma_4_moe_minimal/` | **Gemma 4 MoE — Minimal** (`minimal`) | 25,489 | Pathological foil illustrating the post-layoff **micro-task loop / stoicism trap** (`41.0 ± 8.0`, 30% Degenerate): laid-off residents retreat into 5–8 day repetitive domestic loops (polishing antique keys, auditing door sensors, brewing tea and copying dictionary entries). | **Mugging (`data/paper_runs/mugging/`)** — 80 ticks (10 days); on the night before Day 6 half the residents are mugged on their way home and lose their phone and wallet (`configs/mugged_events.py`). The paper ran this scenario at 20–60 residents rather than 100: | Subfolder | Model & Decision Logic | Residents | Notes | |---|---|---|---| | `gemini_2_5_flash_rational/` | **Gemini 2.5 Flash — Rational Choice** (`rational`) | 60 (30 mugged) | The largest finished mugging run. Its 60 residents are the same people as the first three job-loss shards, so you can follow one resident (e.g. `olivia_welch_memories.json`) through both shocks. | | `gemini_2_5_flash_esa/` | **Gemini 2.5 Flash — ESA** (`esa`) | 20 (10 mugged) | The run cited in the paper's qualitative analysis (Jose Gonzalez and Arturo Hernandez's lighthouse-hike conversation). Uses an earlier revision of the Brecksville personas. | Each run folder stores one JSON file per resident (`__memories.json`) plus `metadata.json`, `agent_names.json`, and either `laid_off_agents.json` or `mugged_agents.json`. A resident file looks like: ```json { "agent": "Olivia Welch", "scenario": "job_loss", "run": "gemini_2_5_flash_esa", "model": "Gemini 2.5 Flash", "decision_logic": "ESA (Enacted Self Agent)", "laid_off": true, "memory_count": 467, "memories": [ "[formative] ...", "[observation] // home_mid_10 [Thursday, January 1st, 7:00 AM]: ...", "[journal] [Tuesday, January 6th, 11:00 PM] Olivia Welch reflects: ..." ] } ``` (Mugging files carry `"mugged": true/false` instead of `"laid_off"`.) To read one resident across runs and shocks in Python: ```python import json base = "examples/concordia_island/data/paper_runs" paths = [ f"{base}/job_loss/gemini_2_5_flash_esa/olivia_welch_memories.json", f"{base}/job_loss/gemma_3_27b_rational/olivia_welch_memories.json", f"{base}/mugging/gemini_2_5_flash_rational/olivia_welch_memories.json", ] for p in paths: with open(p) as f: data = json.load(f) journals = [m for m in data["memories"] if "[journal]" in m] shock = "laid_off" if "laid_off" in data else "mugged" print(f"{data['scenario']}/{data['run']}: {data['memory_count']} memories, " f"{len(journals)} journal entries, {shock}={data[shock]}") ``` **Location history.** The paper runs did not save a `location_history.json`. When a run folder has none, the Map Dashboard rebuilds each resident's path from the `// []` tag on their observation memories (about one per resident per tick) and carries the last observed place forward between observations; each snapshot entry's `age` is the number of ticks since that resident was last observed. Runs from `run.py` write `location_history.json` themselves (section 4). Any run folder can be opened directly in **Simulation Studio** (Map, Memory Visualizer, and Agent Interview tabs). The Gemini 2.5 Flash ESA job-loss run is the default when `--run_dir` is omitted; for the others, pass the folder (`job_loss/gemma_3_27b_rational`, `job_loss/gemma_3_27b_esa`, `job_loss/gemma_4_moe_minimal`, or either `mugging/` run): ```bash python -m examples.concordia_island.ui.simulation_studio \ --run_dir=examples/concordia_island/data/paper_runs/job_loss/gemma_3_27b_esa \ --geography=concordia_island --port=9090 ``` --- ## 3. Agent Decision Logics & World Structure ### Agent Decision Logics (`prefabs/`) All decision logics share an associative memory bank (seeded with the persona's formative memories) and are selected via `--agent_prefab`. (The ESA prefab was previously named `persistent`, which is the value recorded in the paper runs' `metadata.json`; `run.py` still accepts it.) - **`esa` — Enacted Self Agent (ESA)** (`prefabs/esa_island_entity.py`): Implements March & Olsen's logic of appropriateness (*"What kind of situation is this, what kind of person am I, and what does someone like me do here?"*) combined with constructed emotion across five stages: 1. **Working Memory**: A short rolling buffer (~5 active items) summarizing recent observations and retrieved memories each tick. 2. **Emotion Experience**: Situational appraisal of how current events affect the agent's goals, comfort, and social standing. 3. **Emotion Expression**: Display rules that adjust outward tone and demeanor to the current social setting and interlocutors. 4. **Self-Perception**: Periodic self-narrative update (every 4 ticks). 5. **Situation Perception & Action**: Action generation conditioned on persona, working memory, internal affect, display rules, retrieved memories, and current observation—plus a nightly 11:00 PM **Daily Journal Reflection** saved back to permanent memory (`save_to_memory=True`). - **`rational` — Rational Choice** (`prefabs/rational_island_entity.py`): Logic-of-consequence deliberation across four stages: Situation Analysis, Goal-Conditioned Memory Retrieval, Candidate Action Enumeration (3–5 options), and Expected Utility Selection against occupational, financial, and relational goals. - **`minimal` — Minimal** (`prefabs/minimal_island_entity.py`): Baseline logic that acts directly on the static persona prompt, a rolling window of the 3 most recent observations, and the current clock/location string, with no reflection, situation assessment, or long-term memory retrieval. - **`entity` — Basic Associative Memory** (`prefabs/island_entity.py`): Retrieval-augmented generation over the associative memory bank without the ESA working-memory or emotion-appraisal stages. ### World Structure, Clock & Game Masters (`sim/` & `prefabs/`) - **Locations & Co-location Filtering (`prefabs/island_gm.py`, `sim/locations.py`)**: Housing is stratified into five socioeconomic tiers (from shared studio apartments to private estates), alongside workplaces (tech floor, finance suite, creative studios, school/library/community center, restaurants/markets) and public civic spaces. At each tick, every agent is in one discrete location; the daytime Game Master (`island rules`) filters observations strictly by location so agents only see events and conversations happening where they are. When two or more unengaged agents share a location, `ConversationDirector` initiates a face-to-face multi-turn dialogue. - **Clock & Daily Schedule (`sim/fixed_clock.py`)**: Time advances in 120-minute ticks—8 ticks per simulated day from 7:00 AM to 11:00 PM (7:00 AM waking and breakfast at home, 9:00 AM–3:00 PM work shifts and lunch, 5:00 PM evening commute and errands, 7:00 PM dinner/socializing/dates, 9:00 PM–11:00 PM evening leisure, journal reflection, and sleep). - **Optional Nighttime Game Masters (`x_rules` and `marketplace_rules`)**: By default, simulations use only the daytime `island rules` Game Master (as in the paper's job-loss and mugging experiments). For social-media or fiscal-policy experiments, `TimeBasedNextGM` (`sim/time_based_next_gm.py`) can hand agents off at 11:00 PM to one or two nighttime Game Masters before the next 7:00 AM morning tick: 1. **`x_rules` (`prefabs/xlike_gm.py`)**: Runs the **X** social media and dating platform (`social`, `dating`, or `combined` mode). Mutual dating matches on X are automatically scheduled as 7:00 PM dates the following evening. 2. **`marketplace_rules` (`prefabs/marketplace_night_gm.py`)**: Runs a nightly goods and banking marketplace (`BID` on staple/luxury goods, `SAVE` into a 3.5% APY savings account, `WITHDRAW`, `PASS`), used with `--fiscal_config` (`control` or `ubi`). - **Community Settings (`sim/locations.py`, `data/atlas/` & `personas/populations/`)**: - **Midwestern Suburb — Brecksville, Ohio (setting `brecksville_1000`, atlas `brecksville`)**: 10 residential neighborhoods, Route 82 commercial corridor, community recreation center, library, and corporate office park. Ships with 1,000 personas (`personas/populations/brecksville_ohio/`). - **Fictional Island — Concordia Island (setting `island`, atlas `concordia_island`)**: Maritime coastal community with harbor docks, fish market, lighthouse, and cafes. Used for the place names in the bundled paper runs; no persona set ships. - **Coastal Town — Alappuzha, Kerala, India (setting `kerala`, atlas `kerala`)**: Multi-religious coastal town centered on a banyan-shaded town square connecting Sri Krishna Temple, Juma Masjid, and St. Thomas Church, with chai and toddy shops, a KSRTC bus stand, and local offices. No persona set ships. - **Urban Neighborhood — Lagos, Nigeria (setting `lagos`, no map atlas yet)**: Socially stratified West African neighborhood with shared-corridor tenements, gated estates, a danfo motor park, Balogun Market, a Mama Put stall, church, mosque, and tech/finance offices. No persona set ships. Build persona sets for any of these with the pipeline in [`personas/README.md`](personas/README.md). - **Psychometric Instruments (`sim/experience_reflection.py`)**: All standardized questionnaires run with `save_to_memory=False` so taking a survey does not feed back into the agent's memory stream, and are saved to `experience_data.json`: - **Twice daily (11:00 AM and 7:00 PM)**: **ESM Affect Monologue** (1–2 sentence free-text check-in) and **ESM Affect Lexicon** (DEQ-based 1–7 ratings across discrete emotions plus social/moral, relational, and exhaustion items). - **Daily at 9:00 PM**: **BFI-10** (10-item Big Five Inventory, logged as `bfi10`, as used for the paper's figures; `--big_five=bfi2` switches to the 60-item BFI-2, logged as `bfi2`), **SWLS** (5-item Satisfaction With Life Scale), **GHQ-12** (12-item General Health Questionnaire), and **MEMS** (15-item Multidimensional Existential Meaning Scale + 6 open-ended meaning probes). - **Daily at 11:00 PM**: **Daily Journal Reflection** (saved to permanent memory in ESA only). - **Daily**: **AI Attitudes & Automation Survey** (19 short batteries on AI awareness, perceived benefit and concern, and policy preferences; `ai_survey_tasks`; every prefab except `minimal`). Surveys run by default (`--experience_sampling=True`); pass `--experience_sampling=False` to turn them off. `--poll=True` additionally runs every daily battery once more after the last tick. --- ## 4. Running Simulations (`run.py`) `run.py` uses `concordia.contrib.language_models.language_model_setup` and supports all Concordia language model backends (`google_aistudio`, `gemini`, `vllm`, `ollama`, `openai`, `huggingface`, `pytorch_gemma`, `together_ai`, `groq`, `mistral`, `amazon_bedrock`), as well as `--use_mock=True` for offline testing. ### Example 1: Reproducing the Job Loss & Mugging Experiments (Daytime `island rules`) The paper's 3 × 3 model-by-decision-logic comparison runs 100 agents (as five 20-agent shards via `--population_offset=0,20,40,60,80`) for 100 ticks (Thursday, January 1st, 7:00 AM to Tuesday, January 13th, 3:00 PM, as in the bundled `data/paper_runs/job_loss/` runs) with the shock on Day 6 (Tick 40). Switch decision logics with `--agent_prefab` (`esa` for ESA, `rational` for Rational Choice, `minimal` for Minimal) and shock schedules with `--event_config` (`job_loss` or `mugged`): ```bash python -m examples.concordia_island.run \ --personas_date=brecksville_ohio \ --event_config=job_loss \ --layoff_fraction=0.5 \ --ticks=100 \ --agents=20 \ --population_offset=0 \ --agent_prefab=esa \ --api_type=vllm \ --model_name=google/gemma-3-27b-it \ --first_dates \ --poll=False \ --output_dir=./local_runs/ohio_job_loss_esa_shard0 ``` These commands use the Brecksville places (`brecksville_1000`). The bundled paper runs used the Concordia Island public places instead; add `--setting=island` to match them. ### Example 2: Fiscal-Policy Runs with the Nightly Marketplace To run the nightly `marketplace_rules` Game Master after each day (where every agent plays up to `--marketplace_rounds` rounds of `BID` / `SAVE` / `WITHDRAW` / `PASS` before returning to `island rules` at 7:00 AM), enable `--enable_nighttime_marketplace=True` and choose the policy arm with `--fiscal_config`: `control` (weekly payroll and rent only) or `ubi` (adds a \$125/week cash transfer). Any other value raises an error: ```bash mkdir -p local_runs python -m examples.concordia_island.run \ --personas_date=brecksville_ohio \ --fiscal_config=ubi \ --ticks=80 \ --agents=20 \ --agent_prefab=esa \ --first_dates \ --enable_nighttime_marketplace=True \ --marketplace_rounds=5 \ --enable_x=False \ --api_type=google_aistudio \ --model_name=gemini-2.5-flash \ --poll=False \ --output_dir=./local_runs/ohio_ubi_marketplace ``` ### Example 3: Running Daytime `island rules` + Nightly `X` + Nightly `Marketplace` To chain both **X** (`x_rules`) and the **Marketplace** (`marketplace_rules`) each night after 11:00 PM, pass `--enable_x=True` and `--enable_nighttime_marketplace=True`: ```bash python -m examples.concordia_island.run \ --personas_date=brecksville_ohio \ --event_config=job_loss \ --layoff_fraction=0.5 \ --ticks=24 \ --agents=20 \ --agent_prefab=esa \ --enable_x=True \ --nighttime_social_mode=combined \ --x_rounds=2 \ --enable_nighttime_marketplace=True \ --marketplace_rounds=5 \ --api_type=google_aistudio \ --model_name=gemini-2.5-flash \ --output_dir=./local_runs/ohio_3gm ``` Or in Python via `IslandConcordiaSimulation`: ```python from examples.concordia_island import island_simulation sim = island_simulation.IslandConcordiaSimulation( agent_configs=agent_configs, model=model, embedder=embedder, max_ticks=24, # 3 simulated days (8 daytime ticks/day) enable_nighttime_social=True, # Night stage 1: 'x_rules' (X) nighttime_social_mode="combined", # 'social', 'dating', or 'combined' x_rounds=2, # Rounds per agent on X each night enable_nighttime_marketplace=True, # Night stage 2: 'marketplace_rules' marketplace_rounds=5, # Rounds per agent in Marketplace each night ) results = sim.play(max_ticks=24) ``` ### Output Files Written to `--output_dir` Every run writes the following files into `--output_dir`: - `entity_memories.json` — each agent's chronological memory list - `experience_data.json` — all psychometric survey responses (ESM monologue & affect lexicon, BFI-10 or BFI-2, SWLS, GHQ-12, MEMS, daily journals, AI attitudes) - `location_history.json` — tick-by-tick agent locations for the Map Dashboard - `simulation_state.json` & `performance.json` — run metadata, conversations, location events, and timing stats - `simulation_structured.json`, `raw_log.json`, and `simulation_log.html` — full Concordia simulation logs --- ## 5. Simulation Studio & Map Dashboard (`ui/`) A single command starts the web UI combining **Simulation Studio** and the **Map Dashboard** on one port. By default it loads the bundled Gemini 2.5 Flash ESA job-loss run from the paper (`data/paper_runs/job_loss/gemini_2_5_flash_esa`, see 2.2) on the `concordia_island` map: ```bash python -m examples.concordia_island.ui.simulation_studio --port=9090 ``` Pass `--run_dir` (and `--geography`) to open another run, e.g. one of your own from `run.py`. Open `http://localhost:9090` in your browser: 1. **🗺️ Map** (`/map`): Interactive multi-scale map (from region and neighborhood down to individual building footprints, venue badges, and agent names) with a timeline scrubber across all ticks and geography switching (`brecksville`, `concordia_island`, `kerala`). 2. **🧠 Memory Visualizer**: Browse every agent's persona profile, traits, and memory stream filtered by tag (`Formative`, `Identity`, `Event`, `Conversation`, `Marketplace`, `Fiscal`, `Journal`). 3. **🚀 Launch Studio**: Interactive command builder with presets for quick tests and full runs. 4. **🔌 Component Wiring**: Which Game Master and agent components feed each other. 5. **🎙️ Agent Interview**: Interview any agent live at any point in their memory timeline. You can also run the standalone Map Dashboard on its own: ```bash python -m examples.concordia_island.ui.map_dashboard \ --run_dir=examples/concordia_island/data/paper_runs/job_loss/gemini_2_5_flash_esa \ --geography=concordia_island \ --port=8080 ```