> ## Documentation Index
> Fetch the complete documentation index at: https://docs.multivon.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Report browser (view --dir)

> Browse a folder of eval reports locally — a sortable index, any report rendered verbatim, and a diff of two runs with stacked judge reasons. Read-only, no new dependencies.

Once you've run a suite a few times, the reports pile up: one JSON per run,
in a directory, with no easy way to see them together. `multivon-eval view`
already renders a single report as HTML. Point it at a directory instead
and you get an index, click-through, and a diff.

Shipped in [0.15.0](https://github.com/multivon-ai/multivon-eval/blob/main/CHANGELOG.md#0150--2026-06-16).
It runs on the same stdlib HTTP server `view` already uses — read-only,
local, no new dependencies, fully offline.

```bash theme={null}
multivon-eval view --dir runs/           # browse a folder of reports
multivon-eval view runs/                 # same thing — a directory enters directory mode
multivon-eval view --dir runs/ --recursive   # walk subdirectories too (off by default)
```

## INDEX

The landing page is a sortable table of every eval-report JSON in the
directory: suite, model, when, n, pass rate with a Wilson CI bar, error and
flaky badges, and cost. A structural validator decides what counts as a
report — a file needs the real `{summary.pass_rate, cases[]}` shape to make
the table. Anything that doesn't collapses into one "k files skipped" line
rather than being parsed as an empty report. Runs with an error rate at or
above 10% are flagged.

## OPEN

Click any row to open that report at `/r/<idx>`, served by the same
`EvalReport.to_html()` that renders a single file, with a breadcrumb back
to the index. There's no second renderer to drift out of sync.

## DIFF

Pick two runs and `/diff?a=&b=` wraps `report_a.compare(report_b)`:
pass-rate and avg-score deltas, McNemar p with a significance label, and
four buckets — Regressed, Fixed, Still failing, Unchanged. Regressed rows
stack both runs' judge reasons (matched by case input) as prose, so you
read exactly why a verdict flipped instead of guessing.

## Single-file view is unchanged

```bash theme={null}
multivon-eval view results.json
```

A single JSON path still renders one report exactly as before. Only a
directory (or `--dir`) switches into browser mode.