Skip to main content
Directory browsing and comparison are available since 0.18.0. The trial cards, filters, reviewed-candidate workflow and request protections ship in 0.19.0. The viewer serves a self-contained report on loopback. It does not call a model, fetch remote evidence, or write to the report directory. Add --no-browser to print the URL without opening a browser, --port 0 to choose an available port, or --recursive to include reports in subdirectories.

​
Browse and compare

The directory index lists saved suites, models, case counts, pass-rate summaries, errors, variability and costs. A comparison selector opens baseline and proposal side by side. Generated links bind the report path identity and exact file digest: adding a different file cannot silently switch their target. An overwritten report asks you to reload the index. Old bare numeric URLs remain live directory indices; use the generated links when retaining an investigation reference. Comparison follows stable case identity and content checks from compare_reports. Two distinct cases can share an input string without sharing their reasons or outputs. Added/removed cases and comparison evidence issues remain visible. Unavailable significance is shown as unavailable; a non-significant result does not establish equivalence. Legacy ambiguous cases cannot supply reliable pairing.

​
Inspect exact trials

Use the labelled tag/status/search controls to narrow the case list. The summary metrics above it continue to describe the full report. Open Case details with a mouse or keyboard, then open saved trials together to compare repeated runs and attempts. Each trial retains its own output, grader results, errors, origin references, full digest and recorded limitations. Unknown latency and unmeasured checks are labelled accordingly. Report-level capture issues appear before summary scores. A passing score alone is not an acceptance-policy decision. Full saved trial JSON exposes the retained artifact and offers a local JSON download. The HTML contains this evidence, including any retained upstream payloads, so its size and data scope follow the saved report. Legacy aggregate reports cannot reconstruct individual trials that were never recorded. For native Inspect conversations, timing, tool events and media, open the original log in Inspect View. Reference paths are displayed as data; the Multivon viewer does not automatically read or open them. Review collection stays in Label Studio.

​
Turn a reviewed failure into a regression

Download candidate for review creates a candidate bound to the selected trial. It preserves the original case definition and requires review. It never infers the correct answer from the observed output. The failure investigation workflow shows how to reconcile Label Studio reviews, explicitly author the expected result and create a development regression manifest while preserving source grouping.

​
Local access and tested limits

The development viewer checks loopback Host and same-origin browser requests, blocks framing, disables remote content fetching and browser writes, and skips symlinked report files or paths resolving outside the selected directory. These checks reduce accidental exposure and cross-origin access; they are not user accounts or a sandbox against other processes with your filesystem permissions. Keep it local. Standalone exported HTML does not have HTTP response protections. The index, expanded report and comparison were rendered at 1,440- and 390-pixel widths. Native details, labelled inputs, visible keyboard focus and accessible status text support keyboard use. Automated accessibility checks are one part of verification, not a complete screen-reader or accessibility certification. See the validation record.