# Sim Desk

> Set up a finite-horizon multi-server queue - servers, waiting room, exponential mean
> interarrival and service times, horizon, terminating or steady-state analysis with a warm-up,
> base seed, replications and confidence - and find out what the simulation study supports. The
> browser runs an exact port of the SimPy agent skill's queue model and replication runner (same
> seeds, same numbers as replication_runner.py under SimPy 4.1.2). Then a metered reading explains
> what the study supports, or writes the SimPy script that reproduces it.

URL: https://sim-desk.skillsafe.ai/
API: https://sim-desk.skillsafe.ai/api.html
Model: gpt-terra · publisher markup 1000 bps (10%)

## The free engine (in the browser, no account)

- Validates the configuration with the skill's own rules and error messages; accepts the skill's
  replication or queue config JSON.
- Runs every replication: BLAKE2b-derived random.Random streams, the SimPy Resource model,
  time-weighted queue length and utilization, completed-customer waits after the warm-up.
- Reports two-sided Student-t intervals across replications, relative precision and the
  replications needed for a target, totals and unfinished customers, the exact M/M/c/K
  steady-state values as a benchmark, mean queue length in 20 time windows, and paired B minus A
  intervals with common random numbers for a what-if scenario.
- Flags: overload, entity-limit truncation, undefined metrics, event budget, censoring, precision,
  few replications, warm-up drift, benchmark outside the interval, loss systems, short horizons,
  an inconclusive comparison.
- Exports: config.json for replication_runner.py --config, the report in the runner's JSON layout,
  replications CSV, comparison CSV, summary Markdown.
- Checked against SimPy 4.1.2 on 560 random configurations (counts identical, floats to 1e-11).

## The metered lanes (input field `task`)

- `interpret` - a reading of each metric, what the design estimates, the scenario B call (must be
  the browser's), the user's claims judged against the intervals, and what the study cannot show.
- `script` - a SimPy 4.1.2 script that rebuilds the study from the same seeds, checks every
  expected value with math.isclose, applies the fixes and prints the runner config.

Every reply is reconciled on the page: every flag answered, every number found in the browser's
facts or the user's notes, the comparison call, expected values, seed rule, imports and no file
writes checked.

## Sources

- Derived from the agent skill @k-dense-ai/simpy (https://skillsafe.ai/skill/@k-dense-ai/simpy),
  k-dense-ai/scientific-agent-skills by K-Dense Inc. (MIT).
- Kassis et al., Scientific Agent Skills, arXiv:2609.00065 (2026).
- SimPy 4.1.2 (MIT).