Review handoff · offline HTML
Share the evidence, not the editor.
A reviewer should not have to import a project or find the right controls to understand your comparison. Download one read-only HTML file with a navigable case index, source inputs, paired responses and saved review notes.
Try a real editorial handoff
The agent used six public product updates to write a reader-facing brief. The concise A drafts covered the facts but omitted the named next action required by the brief. The action-oriented B drafts supplied one. All six pairs were self-reviewed against four criteria.
Open the downloaded .html file in your browser. It works offline, with no account, installation or remote assets. All content in this sample comes from the public editorial project. Download editable project JSON · Inspect sources and the editorial process.
Not a model benchmark: both drafts and assessments came from the same agent with shared context. B was prepared as a candidate, not as an independent post-hoc experiment. No independent reviewer, held-out sample, human time saving or audience benefit is established. Imported labels and judgments are unverified; this HTML is not signed evidence.
Three questions to answer from the file
- What supports choosing B for this brief? In “Manual criterion patterns,” criterion 3 has six A failures and six B meets on the same six cases. Follow a case link and read the saved note and actual response—not just the count.
- Did the factual content change? Compare the current source input with both recorded responses. The other criterion assessments remain visible. Open the exact recorded prompt to inspect what the writer was asked to do; automatic checks alone do not establish factual accuracy.
- What is not established? This is a selected, self-reviewed six-update batch, not all deployments or a full week. It supports an editorial choice for this brief, not a general claim that B is better for every task or model.
Create your own handoff
- Open your project in the Test Bench. Record responses and review them; pasted drafts alone are not evidence.
- In step 4, choose Download review handoff HTML. Export covers all cases, independent of the editor’s temporary filters.
- Open the file locally. Inspect every case and all expandable details before sharing. Nothing is automatically redacted.
- Send the file through a channel appropriate to its contents. The recipient can open it in a browser without this website. Some mail tools block HTML attachments; use an approved file-sharing channel, or the Markdown report instead.
For printing, expand details you want included first. Project JSON remains the complete editable backup; the handoff cannot be re-imported and cannot collect a recipient’s new judgments.
What the file includes—and leaves out
- Included: all current source inputs and prompts; latest recorded responses with exact prompts and recording-time rules; model/date labels; current check results; saved decisions, criterion snapshots, review notes and earlier review contexts for those latest recordings.
- Project recommendations: all saved choices, rationale, intended use/limits, next-evidence notes and exact decision-evidence snapshots, including sources/cases later changed or deleted. Matching/historical labels compare to the current project. Inspect every snapshot before sharing; deleting a current case does not redact it. See the scoped decision workflow.
- Excluded from current totals: missing or changed-prompt results and out-of-date manual review contexts. Each criterion has its own assessed denominator; paired counts use only the same cases assessed on both variants. No winner is inferred.
- Omitted from the file: draft response/model content and earlier recordings. Their presence/count is disclosed; they remain in the project JSON and full Markdown report. Current prompts and inputs may still contain private draft edits. This is a focused review artifact, not redaction.
- No execution: no scripts, remote styles/images, forms, uploads or tracking. Project text is displayed literally rather than interpreted as HTML or Markdown. Internal links only navigate between sections.
The snapshot does not update when you change the project. A reviewer can inspect your evidence, but cannot verify your claimed model execution or identity from the file.