ITER pack acceptance — 2026-09-29¶

This records real execution on Linux x86-64, not simulated solver or model responses. The pack is part of the unpublished 0.5.2rc3 development preview.

Automatic installation and GUI¶

A fresh deployment directory was created with no runtime. A real Chromium browser clicked Research tools → ITER pack → Install, waited for the numerical readiness report, then clicked Run demo. The application spawned its ordinary installer and monitored simulation jobs. All six checks passed and no browser JavaScript errors were observed. The final repeat took 24.93 seconds with warm package/build caches; this is not a cold-download installation benchmark.

The local runtime occupies approximately 0.9 GB. The demo produces a figure, spectra CSV, IMAS netCDF and JSON. The installer was also run through the CLI, and its numerical preflight executed through Bubblewrap. Repeated installation checks an existing healthy runtime instead of reinstalling it.

On the dedicated remote host, the same pack installed under a non-root account with proot-cooperative, using the managed Python 3.12 fallback because system Python is 3.10. Its numerical preflight and browser-launched demo also passed. This installation lives separately from the active research campaign. PRoot retains its existing cooperative-execution semantics; it is not namespace isolation.

Reproducer: python scripts/verify_iter_pack.py (requires workspace dependencies and Playwright Chromium). Artifacts live under artifacts/iter-pack/, including oneclick-result.json, screenshots, install logs and per-run JSON.

Numerical results¶

Check	Observed result	Qualification
Uniform-sphere ray integral	Maximum absolute chord error 2.5781e-9 m	Analytic geometric integration
CHERAB free-free continuum	Density ×2 produced radiance ×4	Fixed-temperature, fixed-composition scaling
IMAS DD3.39 profile	Exact temperature-array netCDF round-trip	Schema and I/O only
CHERAB-IMAS mock bolometers	Three cameras, five channels each	Mock IDS-to-optics construction
CHERAB-ITER geometry extension	Bézier derivative matched finite difference; Fourier degree convention checked	Compiled helper arithmetic, not a JOREK simulation

No ADAS line-emission tables, ITER internal database or actual instrument CAD was required. Machine-specific applications need separately available data and their own validation. The generated figure is labelled synthetic readiness, not an experimental result.

Real DeepSeek Flash agent acceptance¶

The final run used Simjecture’s built-in agent transport and the provider model ID deepseek-flash, selected for the user’s DeepSeek V4.1 Flash backend. The dated provider revision was not independently verified. It was given the normal skill catalogue, read iter-pack/SKILL.md and the diagnostics reference using read_skill, ran the actual installed capability environment, and authored another CHERAB test.

Completed in 68.73 seconds, within the 480-second bound.
Reported input tokens: 336,440; output tokens: 12,853.
These are sums across provider calls, including repeated context, not unique text or an estimate of charged credits.
A density ratio of 3 produced a radiance ratio of 9.000000000000002.
Its six-point density scan recovered a fitted exponent of 2.0000000000000004.
It retained scripts, plots, JSON and a written report in its project workspace.

This validates task execution and tool use, not novel science or a comparison between model providers. The agent’s prose described the Fourier helper check as “sin 30°”; the actual checked call was py_fourier_mode(30, 1, 2), corresponding to cos(2 × 30°). The numerical test is correct; that prose explanation is not.

Earlier preliminary probes exposed a missing skill manifest: the model fell back to filesystem discovery. That was fixed by shipping manifest.json, then the final test above confirmed direct read_skill use. Preliminary attempts are retained separately and excluded from the final-run token figures.

Solver commissioning¶

JOREK: public revision 4e79ba70cd2a910f9f2a792c9ffb4f074ab8fe90 built locally with GNU Fortran, OpenMPI, MUMPS and HDF5. The bounded model600 example used 256 poloidal elements, n_tor=3, n_period=6 and n_plane=4: an axisymmetric component and an n=6 sine/cosine pair. Two MPI ranks × two threads completed three normalized time steps. The same runtime passed four checks through Bubblewrap: process and step completion, finite HDF5 solution, final normalized time 3 and cross-section area 3.1413801885 m² against π m². It is discoverable as a local registered-build variant; fresh machines still require guided setup. Source/configuration/library copies are preserved in the local .runtime tree.

This is not a converged tearing-mode or disruption calculation. The one-rank attempt failed with MPI_ERR_RANK. Two HDF5 string-size metadata warnings occurred during initialization; all four numbered restart outputs were readable. They are recorded in the report, not suppressed. No remote JOREK solver build is claimed.

DINA-PS: public revision 5c04815895832b55cf13527a86bec7c4ad5f9065 compiled its native core and controllers. Its upstream GREEN component test with supplied ITER configuration completed: 102 passive structures, 12 active coils, 24 loops, 48 probes and 8385 grid points. Full IMAS/iWrap/XMLlib scenario integration was not completed. The GUI keeps DINA as guided setup, not a qualified full solver.

SOLPS-ITER: public revision ca0d04d9b0b2d4887bd3262e4e867d471ba86d62 and its public B2.5/EIRENE/CARRE/DivGeo submodules were inspected. EIRENE standalone and coupled libraries compiled. The complete headless build remained blocked by site library configuration (MSCL/NCARG/NAG-or-BLAS/netCDF) and legacy shell assumptions. The ADAS submodule still references an ITER SSH repository. No coupled divertor simulation or converged SOLPS result is claimed. Its card remains guided setup.

Integration decisions learned from upstream¶

Use supplied cases, native data layouts and documented readers; do not replace a solver when setup fails. Preserve compiled model choices and data dictionary versions. Public source access does not imply access to all machine data or atomic databases. Skills route to focused upstream material and record tested pitfalls; optional helpers do not prescribe an entire research strategy.

IMAS-Codex is an existing schema-oriented MCP interface, separate from these solver workflows. This pack uses the local IMAS APIs and skill interface; it does not silently connect an agent to a hosted service or claim an upstream solver agent interface that was not found.

Repository checks¶

The full suite passed: 748 passed, 7 skipped in 181.44 seconds. Skips cover optional unavailable runtimes and local PRoot requirements; the remote PRoot pack demo was exercised separately. Changed Python files pass Ruff, JavaScript syntax checks pass, and documentation builds with warnings treated as errors. The skill frontmatter validator also passes.