# Core-Hour Check

> A planning desk for bounded fluidsim 0.9.0 computational-fluid-dynamics runs at https://core-hour-check.skillsafe.ai/. Paste or drop a strict JSON run plan; the browser runs JavaScript ports of the three standard-library planning tools bundled with the fluidsim agent skill - solver_config_validator.py, grid_resource_estimator.py and simulation_dry_run.py - with stdout and exit codes identical to the Python, for free. A metered run (model gpt-terra) then drafts a strict plan from a description of the study, or reviews a plan against the skill's scientific acceptance gate. Derived from the agent skill @k-dense-ai/fluidsim (k-dense-ai/scientific-agent-skills).

## What the free lane computes (in the browser, nothing uploaded)

- solver_config_validator.py: the plan against the skill's static fluidsim 0.9 profile of the Cartesian pseudospectral solvers (ns2d, ns2d.bouss, ns2d.strat, ns3d, ns3d.bouss, ns3d.strat, plate2d, sw1l and variants, waves2d) - unknown keys (including the old CFL), grid sizes and lengths, explicit dealiasing (warning above 2/3), explicit time bounds and scheme, forcing, output flags and save periods, safe output and restart paths, the nine scientific statements (boundaries must be periodic), resources (no CPU oversubscription), serial versus mpi-preview mode, pinned fluidsim 0.9.0 and fluidfft 0.4.5. It never establishes numerical convergence or physical validity.
- grid_resource_estimator.py: resident state from the grid and the solver's real/complex field counts, peak memory (x workspace 8 x safety 1.5 by default), state snapshots, spectra, budget and scalar output storage, and output-file count, each against the plan's ram_gib, disk_gib and max_output_files. Runtime is never estimated.
- simulation_dry_run.py: the launch script (dry run by default; --execute needs the exact reviewed config id), its SHA-256, the approval command and, for mpi-preview, the launcher preview. Nothing is launched.
- Page facts: grid spacing dx = L/n, k_max after truncation = coef_dealiasing x pi / dx, minimum step count t_end / deltat_max, numbered flags for every validator error and warning, resource overrun, forbidden label (DNS, converged, validated, steady, physically correct) and placeholder provenance id, and a browser verdict (ready_for_pilot / revise_first / not_plannable).
- Checked against CPython 3.11.2 on 16,200 runs of fuzzed plans (15,000 also under 3.13.6): stdout and exit code identical in every run, including 290 runs where the Python stops with an uncaught exception (the page reports those as crashes). When two or more scientific text fields are invalid the Python names one chosen by its hash seed; the page names the first in listing order and lists the rest.
- Exports: the plan, the launch script, the three tools' reports as JSON, a Markdown summary and the equivalent command lines.

## The metered lanes (task field)

- draft: a complete strict plan sized as a tiny serial pilot, every chosen value listed as an assumption, open questions and (for a revision) the changes. The page re-runs the skill's three tools on the drafted plan and reports what they say.
- review: every flag answered (confirmed / dismissed / explained), the twelve gate items (units, boundaries, initial_conditions, forcing, resolution, timestep, budgets, refinement, benchmark, acceptance, resources, provenance) rated addressed / thin / missing with evidence, concerns, pilot checks, questions and what is not established; verdict never looser than the browser's.

Input: {task, plan?, facts? (JSON string built by the page), brief? (draft), notes?, question?}. The model never runs anything and is told never to compute a new number; the page reconciles every reply. API: https://core-hour-check.skillsafe.ai/api.html

## Limits

Only the plan schema the skill's validator covers (Cartesian pseudospectral solver profiles). The HDF5-reading tools of the skill (output_inventory.py, budget_summary.py, restart_compatibility.py) are not ported. A passing plan establishes neither numerical convergence nor physical validity.

## Source

Agent skill: https://skillsafe.ai/skill/@k-dense-ai/fluidsim (repository https://github.com/k-dense-ai/scientific-agent-skills, skills/fluidsim, MIT). fluidsim paper: Mohanan et al. 2019 ("FluidSim: Modular, Object-Oriented Python Package for High-Performance CFD Simulations"), Journal of Open Research Software 7(1) 14, doi:10.5334/jors.239; FluidFFT: doi:10.5334/jors.238. No fluidsim code is included; see https://core-hour-check.skillsafe.ai/NOTICE.txt

Core-Hour Check is independent and not affiliated with the FluidDyn project (fluidsim) or K-Dense Inc.