0.5.3rc3 scope and validation

This preview publishes the maintained scientific coding-agent benchmark and improves its public presentation. It retains the local/SSH experiment workers from rc2.

User-facing changes

  • Two official tasks: radiation energy accounting and RZ plasma diagnostics.

  • Actual owned measurements: 42 scheduled configurations, 252 scheduled attempts, 228 graded attempts and 24 observations without inference. Public task tables contain only configurations that reached inference: 40 RZ rows and 41 CSV rows.

  • Independently ranked verified finish-time bars, per-attempt cost bars and a cost–time Pareto plot with the upper-left corner preferred.

  • Distinct model colours shared across plots and a legend including single-effort models. Dashed lines join efforts of the same model and coding agent. Frontier points are outlined individually; no line joins different models.

  • Unfinished runs appear in red and retain their recorded cost bars. Partial receipts show a labelled lower bound and receive no cost rank. Configurations without inference are omitted from public rankings.

  • Remove the completion/time plot, pilot and unranked panels. Methodology is a documentation link beside the plot, rather than a dropdown.

  • Current custom-model/community submissions have their own view. Older qualifications and owned grades remain available in audit downloads; importing an owned grade does not turn it into a new community measurement.

Scientific boundary

These tasks analyze recorded plasma diagnostics. They do not launch new FLASH or WarpX simulations, qualify a physical hypothesis, or establish general research ability. RZ has one measured attempt per configuration. The five CSV repetitions measure delivery on the same task, not generalization across independent problems.

Costs use published API-equivalent rates and recorded usage, rather than subscription invoices. AGY counters require an explicitly labelled interpretation; interrupted streams may provide only lower bounds. Host grades and scientific scores were not changed by the presentation fixes.

Validation

The release checks cover the full Python suite, DSH/MCP adapter, lint, strict docs, built distributions and a fresh-home installation of the workspace bundle. The installer test supplies no API key or CLI agent and changes no host system packages.

Browser validation covers both task tabs, model/effort filters, cache cost bases, unfinished spending, absent no-inference rows, individual frontier markers, the methodology link and mobile overflow. Both the live workspace and standalone publication were checked without JavaScript errors.

The bilingual benchmark videos use stock synthetic narration and independently validated 1080p media/captions. They are production artifacts, not benchmark evidence or required runtime dependencies. No extra model inference was performed to make or revise them.