task_id stringclasses 10
values | title stringclasses 10
values | description stringclasses 10
values | instruction stringclasses 10
values | task_path stringclasses 10
values | scored_outputs int64 5 5 |
|---|---|---|---|---|---|
meat-bro | Meat Bro | Predict the combined effects of pressure, hormone changes, arterial stiffness, and liver impairment. | # Meat Bro under pressure
An adult male, nicknamed Meat Bro, is obese and has 50% more skeletal muscle
than the reference adult male. He eats a normal diet. Both subjects have the
same body composition, diet and 24-hour preparation before comparison time zero.
Apply these changes together to the intervention subject ... | tasks/meat-bro | 5 |
scenario-02 | Summit Air Scrubber | Predict the combined effects of altitude, exercise, and altered breathing gas. | # Scenario 02
An adult male begins awake,
resting and lying down at sea level in still air at 72°F (22.22°C), 30% relative
humidity, wearing normal clothing.
- At 0.1 min, the intervention subject sits and moves to an altitude of 1,000 m.
- At 5.1 min, altitude becomes 2,000 m; at 10.1 min, it becomes 3,000 m.
- At 1... | tasks/scenario-02 | 5 |
scenario-03 | Rescue Basket | Predict responses to blood loss, cold, and head-up tilt. | # Scenario 03
An adult male begins awake,
resting and lying down at sea level in still air at 72°F (22.22°C), 30% relative
humidity, wearing normal clothing.
- From 0.1 to 5.1 min, the intervention subject loses whole blood at a fixed
80 mL/min, totaling 400 mL. Bleeding is then stopped. No replacement fluid is giv... | tasks/scenario-03 | 5 |
scenario-07 | Heart with a Shortcut | Predict exercise and recovery responses with weakened heart function and a vascular shunt. | # Scenario 07
An adult male has a shared
120-minute preparation before comparison time zero:
- Both ventricular basic-contractility settings are set to 0.70 times their
reference values and remain at that setting throughout.
- During the first 30 min, whole blood is removed at 40 mL/min while an equal
40 mL/min o... | tasks/scenario-07 | 5 |
scenario-09 | Robot Paramedic | Predict responses to saline and mechanical ventilation after blood loss with impaired kidney function. | # Scenario 09
An adult male has a shared
150-minute preparation before comparison time zero:
- At the beginning of preparation, each kidney's filtering-nephron percentage
is set to 50%, with total nephron counts unchanged. This setting is retained.
- After 120 min, whole blood is removed at 50 mL/min for 15 min, to... | tasks/scenario-09 | 5 |
scenario-11 | Shunt Shutdown | Predict the effects of closing a vascular shunt while changing adrenergic signaling and oxygen supply. | # Scenario 11
An adult male undergoes a shared 120-minute preparation. Both ventricular basic-contractility settings are 0.70 times their reference values. During the first 30 minutes, whole blood is removed at 40 mL/min and replaced with the same volume of plasma containing no red cells, totaling 1,200 mL exchanged. ... | tasks/scenario-11 | 5 |
scenario-12 | Oxygen-Loop Fault | Predict hormonal and metabolic responses to mechanical ventilation with elevated oxygen and carbon dioxide. | # The Oxygen-Loop Fault
An adult man at rest lies supine at sea level. For one hour, a breathing circuit supplies dry gas containing 60% oxygen, 1% carbon dioxide and 39% nitrogen. Mechanical ventilation delivers 8 breaths per minute with a tidal volume of 700 mL. Room temperature is 22.2 °C. Compare with the same man... | tasks/scenario-12 | 5 |
scenario-13 | Centrifuge Saline Bag | Predict the combined effects of increased gravity, head-up tilt, supplemental oxygen, and saline. | # The Centrifuge Saline Bag
An adult man at rest spends one hour restrained at 55 degrees head-up in a centrifuge producing an effective gravity of 1.5 g. He breathes dry 40% oxygen and 60% nitrogen at sea-level pressure and receives 500 mL of 0.9% saline evenly over the hour. Room temperature is 22.2 °C. Compare with... | tasks/scenario-13 | 5 |
scenario-16 | Sealed Ambulance | Predict responses to concentrated saline and altered ventilation after blood loss with impaired kidney function. | # The Sealed Ambulance
An adult man has half the usual number of filtering nephrons in each kidney. After two hours in this state, he loses 750 mL of whole blood evenly over 15 minutes and rests for 15 minutes after bleeding stops. He then spends one hour supine breathing dry 40% oxygen, 1% carbon dioxide and 59% nitr... | tasks/scenario-16 | 5 |
scenario-17 | Cold Transfer | Predict the combined effects of cooling, head-up tilt, and supplemental oxygen. | # The Cold Transfer
An adult man at rest starts supine in a 22.2 °C room, breathing ordinary air at sea-level pressure. At time zero, he is moved into a 15 °C room, restrained at 45 degrees head-up, and supplied dry 60% oxygen and 40% nitrogen. He remains under those conditions for one hour, without exercise, bleeding... | tasks/scenario-17 | 5 |
Physim Reasoning Tasks
Ten whole-body physiology tasks. Each task combines interventions across organ systems and asks an agent to predict five downstream changes relative to a matched simulated control.
The task catalog, model comparison, and project overview are available at symbolic.life/tasks.
Contents
tasks/: runnable Harbor tasks, each with an instruction, container environment, verifier, and reference solution.task_catalog.jsonl: browsable task descriptions and prompts.download/: the exact task archive served on symbolic.life.benchmark-results.json: completed six-model benchmark results.
Run
Install Docker and uv, then run from this repository’s directory:
uv run --with harbor==0.23.0 python run_verified.py run --path ./tasks --agent oracle
To evaluate an agent, configure its provider credentials in your own environment and run:
uv run --with harbor==0.23.0 python run_verified.py run --path ./tasks --agent terminus-2 --model PROVIDER/MODEL --ak reasoning_effort=high
The supplied launcher uses the tested Harbor version and mounts each task’s verifier separately and read-only. Keep tests/ and solution/ outside the evaluated agent’s environment.
Scoring
The answer gives increase, decrease, or unchanged for each of five outputs. The task prompt specifies the comparison and unchanged threshold. Output accuracy is the fraction of correct directions; whole-task success requires all five directions to be correct.
Results
Each of the six models completed three attempts per task. The model identifiers and reasoning settings are recorded in benchmark-results.json. All recorded scores were independently regraded against the released task files.
| Model | Correct outputs | Output accuracy | Fully correct attempts |
|---|---|---|---|
| Claude Opus 5 | 73/150 | 48.7% | 0/30 |
| GPT-6 Astra | 69/150 | 46.0% | 0/30 |
| Nemotron 3 Ultra | 68/150 | 45.3% | 0/30 |
| Grok 4.6 | 66/150 | 44.0% | 0/30 |
| Gemini 3.8 Flash | 63/150 | 42.0% | 0/30 |
| Grok 4.7 | 60/150 | 40.0% | 0/30 |
Data source
Reference answers are frozen outputs of a physiological simulation. They are model predictions, rather than measured human outcomes. Running and grading the released tasks does not require access to the simulator or a GPU.
License
The task data, prompts, and included code are released under the Apache License 2.0.
- Downloads last month
- 36