Review, report, and share results

When the run reaches the Review the answers phase, the benchmark app shows two answers for the first prompt. It does not tell you which context arm produced which answer, so your choice cannot be influenced by the label.

For each prompt:

Read Answer A and Answer B.
Answer the required question, Which answer would you rather use for this work? Choose Answer A, Answer B, Both are good, or Neither is good enough.
Optionally answer Which answer is more accurate and grounded in the available context?
Optionally add a note about what made the difference. Notes are the most useful thing you can give Atlassian if you later share the report.
Select Next to move to the next prompt. Your choices save as you go, and Previous takes you back.

After you choose, Show RelIQ hint appears. It reveals what the automated judge preferred and why. It is deliberately hidden until you have committed to an answer, so leave it alone until then.

Select Submit review once every prompt has a choice.

Success: the phase changes to Building the final report, then to Benchmark complete, and the report opens in the same tab.

Judge the answers on whether you could act on them. An answer that is confident but wrong about your work is worse than a shorter answer that is correct and says what it does not know.

Read the report

The report header identifies the run, the agent, the model, and when it was generated. Two badges matter: the privacy classification of the evidence, and whether the run is Sharing eligible.

Section	What it tells you
Setup	What was held fixed and what differed between the two arms. Read this first if any number looks surprising.
Prompt outcomes	One row per prompt, showing which arm you preferred, the automated quality result, and token and error counts for each arm. Expand a row to see both answers, the judge rationale, and the tool traces.
Technical analysis	How quality was scored, and the aggregate comparison across prompts.
Method and limits	What this run can and cannot support.
Report metadata	Run identifiers and generation details.
Actions	Sharing eligibility and suggested next steps.

Four results carry most of the meaning:

Result	How to read it
Preferred answer	Which answer you chose in the blind review, shown per prompt as preferring Teamwork Graph context or Siloed context. This is the main quality result.
Automated quality	The RelIQ score, from an LLM judge that compared the answers without knowing which arm produced them. Supporting evidence, not the verdict.
Token use	How much text each arm processed and generated.
Time	How long each arm took.

Read quality and token use together. Fewer tokens is only a win if quality held. A cheaper answer that dropped content is not a saving, it is a shorter answer. The results worth acting on are prompts where Teamwork Graph context is at least as good on quality and cheaper on tokens, or clearly better on quality at a similar cost.

One run is noisy. Run the benchmark more than once before drawing a conclusion. Results move with model variation, how fresh your connected sources are, and which details you put into each prompt.

If some prompts could not complete, the report says so with Evidence is partial and states how many of the planned comparisons it contains. Treat a partial report as a partial result.

Everything stays on your computer. Results live under your benchmark home directory, ~/.benchmark by default, and nothing is sent to Atlassian unless you package a report and send it yourself.

If the report offers Open static report, that export is a self-contained local HTML copy you can keep or read later.

If you ran the benchmark in the app, package that session from the terminal:

1
2
benchmark share browser

The command lists completed browser sessions that have value.html and value.json under report/, post-run/pair.json, and a completeness of complete, partial, or unscored. It redacts the HTML report and opens a local review page. Download the archive from that page, then unzip it before opening Technical or side-by-side links. The first successful package for a session is locked: if that archive is deleted, share the session again only after you restore the exact file or run a new session. From a script or other non-interactive terminal, pass --session and review the archive later from a real terminal.

To package a completed terminal run, use:

1
2
benchmark report zip

To package a specific run:

1
2
benchmark report zip --run-id 20260728T120939Z-a1b2c3d4e5f6

The bundle contains the run's redacted JSON evidence: results, answers, traces, payloads, logs, lifecycle state, and the available score, automated review, and blind-review records. It does not contain generated HTML, because reports are rebuilt from the evidence when the bundle is opened.

Open a bundle again with:

1
2
benchmark report unzip --report-name 20260728T120939Z-a1b2c3d4e5f6-evidence.zip.gz

The bundle stays local, under your saved runs folder such as ~/.benchmark/saved_runs, until you send it through the channel you agreed with Atlassian. Review the contents before you share it.

If you ran the benchmark from an AI agent, you can also ask the agent to package the latest run for you. It uses the same redacted bundle.

Next step

Run it again with different prompts, or see Troubleshooting if something did not work.