0.5.4rc1 scope, checks and limits

This preview adds execution strategy control to minimal research. The worker chooses its scientific approach; an independent director can stop named experiments and request a recorded replacement plan or reasoned challenge. Methods approval and scientific claim acceptance remain separate reviews.

Included behavior

  • A default-on director for new minimal studies, with visible GUI launch settings, model/effort route, decisions, stop confirmations and worker responses. CLI users can select --director or --no-director. Existing studies retain their saved policy.

  • Strategy checks during long experiment waits and sustained planning, bounded director calls, named cancellation and available partial-artifact retention. A later continue review cannot waive the latest unacknowledged replan. A failed stop control remains visible without losing the valid directive or preventing other requested stops. Recovered reviews clear stale review errors.

  • Optional bounded JSON/FLASH live timing through local and SSH execution, receipt-backed advisory progress, measured cost and conditional runtime estimates. Older workers remain usable without monitoring; advertised support is required when a new experiment requests it.

  • Reserved-slot MPI launch helpers, read-only compiler alternatives and dependency source manifests for OverlayFS hosts with unstable inode numbers. Explicit trusted process-cooperative execution is selectable in CLI/TUI/browser machine settings.

  • Frozen exploratory analysis, clearer method experiment-reference validation, separate descriptive limitations and native reviewer read-only tool events.

  • GUI discovery of externally launched studies without a server restart.

Verification

The four new recovery regressions and the existing director suite pass: 18 tests. A full local pass completed 1,166 passing tests and seven optional runtime skips, and exposed one legacy tool-free reviewer compatibility failure. The permission was made explicit for read-only minimal research contexts; the original legacy guard remains intact. All 82 affected supervisor, director, journal, oversight and browser tests then passed, including the original failing case. The exact release source receives the full CI gate before publication.

All 14 DSH/MCP tests, Ruff on the clean tracked source, JavaScript syntax and strict Sphinx passed. The desktop/mobile monitor tests show failed stop-control messages without errors or horizontal overflow. The source distribution, wheel and workspace bundle exclude local environments, credentials and untracked research data; their modified runtime files match the source bytes. A checksummed bundle installed in a fresh HOME with no model credentials or native agent and no system-package changes. Its GUI, assets, HDF5 reader and packaged skills passed the installer verification. The release also retains the earlier execution qualification and director/adaptive commissioning report, which distinguish launcher readiness from scientific accuracy.

The ongoing 12-hour Al campaign has completed unperturbed and coherently perturbed 48×48×24 Cartesian FLASH trajectories through 40 ns, each in approximately 51 minutes. The director stopped a larger, poorly specified stochastic follow-up and the worker implemented a replacement. These are operational observations from an unfinished campaign, not a benchmark of scientific completion or model cost. The candidate’s additional recovery changes are verified separately; this running campaign continues on its original private deployment.

Limits

  • The campaign has no accepted methods or reconnection attribution result at the recorded checkpoint. Spatial resolution, temporal accuracy, energy accounting and kinetic-reference uncertainty still require qualification. Small changes in radiation output between perturbations do not measure direct reconnection energy.

  • The broad adaptive temporal screen failed despite completing all commissioned trajectories. Its failed results remain preserved. FLASH guidance prefers qualified adaptive stepping and affordable complete exploration, not arbitrarily large forced timesteps.

  • Director calls can time out and add substantial context/tool overhead. Native usage counters distinguish worker/director roles and cache usage, but do not supply provider invoices or complete per-request counts. This preview does not claim reduced spending or guarantee useful scientific judgment.

  • Linear completion forecasts can be misleading during startup ramps or changing plasma conditions. They are mutable operational telemetry, not evidence.

  • Cooperative process and PRoot execution are not OS security boundaries. Native agent host tools remain available. Source manifests verify declared bytes; they do not qualify the solver’s numerical method.

  • No new model leaderboard edition, full WarpX patch investigation or experimental reproduction is claimed. Structured and frontier supervision retain their existing behavior.