Production: Fuller Card Previews, Public Results API, and BYOK Costs

Deploys main at 6efe20f.

• #304 Prompt hovered cards' readers to open the page, fill the preview, and open repository pages side by side
• #303 Add Mintlify docs and public results API
• #301 Count BYOK provider charges in gateway generation costs

Production: Confidence Intervals on Results Charts and Verified Task Images

Deploys main at 3057a78.

• #298 Show 95% confidence intervals on the results charts
• #276 Start trials from verified task images
• #296 Use gateway-reported charges for evaluation costs when available
• #297 Move one-off operator tools out of live source
• #295 Simplify Results and empty states

Production: Results Charts Without Error Bars and E2B Plan Errors

Deploys main at 1d033ba.

• #302 Remove confidence-interval error bars from the results charts
• #300 Say when an E2B key's plan is too small for 24-hour sandboxes

Production: Temporal Workers on GKE and Dari Routing for Evaluations

Deploys main at e61843e.

• #293 Run every Temporal worker on GKE, on nodes system pods can't preempt
• #294 Allow Dari Routing for production evaluations
• #290 Keep the overlapping-lists test from depending on how fast CI starts requests
• #288 Set up gcloud once in prod deploys too

Production: Browsable, Downloadable Tasks and a Faster Results Page

Deploys main at ba70ba1. It also carries #279, whose own release (prod-20261002-20b1479) was cancelled before deploying, so one deploy ships both.

• #280 Publish a release's tasks on selfbench.dev, to browse and download: Publish Tasks in the release dialog; a Tasks list and viewer on repository pages; downloads without the repository snapshot; task data kept out of search and training crawls, cached a minute, and marked with a canary once its secret is set (Terraform adds the empty selfbench-task-canary secret)
• #279 Keep run summaries in the database so the Results page stops reading every run's record

Production: Time-Limited Trials Scored and Backfillable

Deploys main at 6e6fd60. - #282 Score trials whose agent hit its time limit instead of dropping them: Harbor's verifier result now counts, Pi timeout costs are recovered, timed-out trials are marked in the app, and backfill-agent-timeouts.js can score earlier timeouts

Production: Long-Prompt Costs Priced Where Rates Change

Deploys main at 2a70a03. - #284 Price long prompts only where rates change, and recover missing eval costs - #287 Find Claude Code's own cost by its result event - #285 Close the task viewer on a click around it, and make the snapshot rebuild one command - #286 Publish a release's tasks by default, and box Publish Tasks so it is plain to see

Production: Faster Results Page from Stored Run Summaries

Deploys main at 20b1479.

• #279 Keep run summaries in the database so the Results page stops reading every run's record (adds the evaluation_summaries table)
• #277 Open the README with a selfbench.dev demo and results chart (README only)
• #278 Explain how SelfBench works with an animated GIF (README only)

Production: Canary-First Task Instructions and Run Summaries Stored Under Load

Deploys main at 019c361.

• #283 Open each published instruction with the canary, and show it without
• #281 Store the summaries a list of runs rebuilds, however busy the server is

Production: Results Grouped by Configuration and Repository Page Headings

Deploys main at ee0918a.

• #236 Group the Results page by configuration, run and task
• #275 Make a repository page's name its h1