0.6.0 release acceptance

Version 0.6.0 introduces the hosted service and a shared local/hosted chat layout. It includes the research director, recovery and simulation provenance work from the 0.5.4 release candidates. The scientific claim rules remain unchanged.

Interface and account checks

Browser checks cover compact/long/cleared drafts, editor expansion, the model popover, theme persistence, account/workspace details, responsive panel resizing, and existing conversation/study/evidence navigation. A live HTTPS check at simjecture.com found no JavaScript errors or horizontal overflow at 390 px. The rail stays dark in both themes; its monitor/account controls sit at the bottom. Inspector overflow controls and mouse-wheel navigation keep all tabs reachable. The home suggestions name actual Python, FLASH and autonomous hypothesis tasks. Interactive and Autonomous research have brief descriptions; research views use peer navigation tabs. Hosted benchmark checks cover both tasks, all three charts, result downloads and disabled testing/import controls. The hosted API rejects benchmark mutations and exposes shipped public results without tenant files.

Hosted API checks cover account-scoped files/jobs, quota reservations/refunds, OAuth state/PKCE/verified identity, owner authorization by immutable GitHub ID, memory-only credentials, file-upload ownership and default attachment downloads. Deletion checks cover ownership, the same account on another device, explicit confirmation, stopped writers, restart recovery, erased content and retained usage counters. A real local child process is stopped before its folder is erased. Guest cleanup tests cover last-tab expiry, another live tab, refresh/sign-in protection, cancellation/erasure and signed-in persistence. Brief/cutoff checks freeze the queued requirements and preserve a shorter positive execution budget.

Actual numerical and execution qualification

The dedicated Linux/P40 host executed FLASH Sod 1D, Orszag–Tang MHD 2D and Sedov RZ examples, WarpX CUDA Langmuir examples in 1D/2D/RZ/3D, and the installed ITER diagnostics/data demonstration. Each used the recorded experiment launcher. These are numerical readiness cases with declared model scopes.

The same confined launcher also executed the upstream CUDA bremsstrahlung and inverse-bremsstrahlung collision cases for 100 steps. Both produced photon energy; the recorded electron/ion/photon energy sums drifted by less than 1e-10. The installed emission tables cover H, He, B and C, not aluminium. These checks qualify collision operators, not a full Z-pinch implosion or radiation-yield model.

The general executor additionally passed private-file denial, custom C compilation and executable reuse, editable WarpX CUDA and FLASH inputs, and compilation/run of FLASH’s resistive-MHD island-coalescence application from a private writable copy of the read-only source. The compilation took about 26 seconds and the 64² run about 3 seconds on that host. A live console-path replacement test verified that the broker reads its original descriptors rather than a job-controlled path. Unit checks also reject unsafe paths, symbolic/hard links and binary console dumps.

A real MiMo v2.6 Flash chat task integrated a double pendulum with SciPy and produced trajectory/energy plots, CSVs and a summary. The completed acceptance turn used 9 requests, 76,793 reported input tokens and 5,593 output tokens. Its reported energy drift was checked within the numerical example’s own scope; this interactive result did not undergo autonomous scientific claim review.

Testing exposed and repaired a FLASH-helper assumption about the native_ filename prefix, incomplete raw-output archives for other application names, and ineffective follow-up access to native artifacts. Snapshot selection now uses recorded physical time, and native outputs have project-relative analysis paths. A failed/limited agent attempt remains a failed/limited record; installation or a successful numerical run is not counted as a completed scientific task.

A completed owner task also ran the 64² FLASH island-coalescence preset to t=0.05 and produced field/energy plots through custom Python analysis. Its first plotting call failed because the output directory was absent; a fresh call created the directory and saved the plots. This exposed a presentation bug that classified analysis commands with declared outputs as simulations. Execution categories are now explicit, and old untyped custom commands are projected into Commands without changing their scientific receipts. Hosted command outputs retain their own provenance and image previews.

A further live MiMo task integrated a harmonic oscillator and ran diagnostics separately. The completed turn used 6 requests, 36,853 reported input tokens and 2,248 output tokens. Its solver run appeared in Simulations; CSV inspection and plotting appeared in Commands. This check exercises the model-facing execution category through the actual confined launcher and hosted workspace projection.

Limits

The optional general execution profile uses separate UIDs, Landlock and seccomp, with CPU/memory/process/storage/time bounds. It requires Linux enforcement support and an operator-owned broker; it fails closed and requires a private Unix HTTP gateway. It does not provide a network namespace or establish a complete external security audit. Operators must protect other network services and respect solver and data licences. CUDA uses device permissions and a shared pool rather than hardware-level tenant partitioning.

Shared admission/rate/resource limits and provider credit still apply to owners. Application budgets are not provider invoice ceilings. Guest cleanup has an explicit refresh/network grace period; minimal allowance counters are retained. Signed-in retention/backups remain operator responsibilities. The self-hosted workspace does not acquire these hosted retention rules.

These checks demonstrate software behavior and bounded numerical execution. They do not establish reconnection rates, radiation excess attribution, machine validation, cross-model superiority or lower inference cost.