How do I share an Inspect AI eval log with someone who doesn't run Inspect?

Run `inspect view bundle --log-dir logs --output-dir dist`, which writes a self-contained static viewer, then publish that directory to Comma. Reviewers open a URL and get the sample browser, the transcripts and the scores — with no Python environment.

Why can't I just open the bundled viewer's index.html?

The viewer fetches the log files and its own JS as modules, and browsers block those fetches from `file://` origins. It has to be served. Publishing serves it once, for everyone, at a permanent URL.

Can I share a single .eval log instead of the whole directory?

Bundle it — point `--log-dir` at a directory containing just that log. The bundle carries the viewer plus the logs, so there is nothing for the reader to install or point a tool at.

Is it safe to publish eval transcripts?

That depends entirely on what is in them, and eval transcripts frequently contain model outputs you would not want indexed. Reports are private by default; you choose per report whether it stays private, goes to your team, is gated to your email domain, or is opened to named reviewers.

Can reviewers comment on a specific sample?

Yes. Threads anchor to the rendered content, so 'the grader is wrong on this one' sits on that sample instead of in a channel, and survives into the next bundle you publish.

Share an Inspect AI eval log

Inspect writes .eval logs and gives you inspect view to read them — a local server, on your machine, pointed at your log directory. Which is fine until the interesting result is the one somebody else needs to look at.

The escape hatch is already in the CLI:

inspect view bundle --log-dir logs --output-dir dist

That writes a self-contained static viewer: the log viewer app plus the logs it reads. Everything a reviewer needs, in one directory, and still unopenable by double-clicking — the viewer fetches its logs as modules, and file:// origins block that.

Publish the bundle

Drag dist/ (or a zip of it) into the app:

index.html becomes the report body.
The viewer bundle and every log file upload as assets, with their relative references rewritten to the uploaded copies.
Scripts run inside a sandboxed iframe (allow-scripts, no allow-same-origin), so the sample list, the transcript panes and the score filters all work.

From a pipeline, zip it and POST/PATCH through the API against a stable report id so each eval run appends a revision instead of scattering links.

Keep the transcripts private on purpose

Eval logs are not marketing assets. They contain full model transcripts, often on prompts you would rather not have indexed. Every report is private until you say otherwise; from there the choices are team, email-domain gate, named reviewers, or a public link — see the sharing model and, for a link you want gated by a password rather than an account, password-protecting a report.

Review is the point

The disagreement in an eval review is almost never about the headline score. It is about individual samples:

Anchored threads on the sample whose grade is wrong — see commenting on HTML.
Revisions — one report id per eval, one revision per run, so a scorer change is a diff rather than two tabs and a squint.
Agent loop — the agent that runs the eval can read those comments back through MCP and adjust the task or the grader.

Limits

Entry HTML: 5 MB. Assets: 25 MB per file, 250 MB and 500 files total, 500 files max. A bundle over many hundreds of samples can hit the file count — bundle a filtered log directory instead of everything.
60 requests/minute per token.

Try it

Comma is free — unlimited reports, unlimited commenters, unlimited revision history.

Publish an eval bundle →