How to run a literature review

You will need:

• Your PubMed search queries (the same ones from the previous review packet)
• The date range of the update window
• An idea of the maximum number of articles you want to carry into the evidence table — the workspace defaults to a sensible cap based on your prior packet

Sign in with your Chemular email at the sign-in page. You'll get a 6-digit code in your inbox; type it back into the page. There is no password.

Click Include for every record you want to keep. That's it. Records you don't Include are out of scope — you do not need to mark them excluded, click anything else, or justify leaving them. If you walk away from a record without clicking Include, the system treats that as “the scientist chose not to include this study,” which is exactly correct.

This applies at every screening stage: title-and-abstract review and full-text review. Include is the only verb you owe the workspace.

From the home page (the page you land on after sign-in), click New projectif this is a new product or a new review topic, or open an existing project if you're running an update.

Inside a project, click New run. A run is one execution of the review — one update window, one set of search queries. Each run produces its own deliverables and never overwrites a previous run.

Fill in the run setup form: the review title, the update window (start and end dates), and your PubMed query strings. Save. You'll land on the run dashboard.

On the run dashboard, the primary blue button at the top right says Run everything →. Click it. The workspace will redirect you to a progress page.

The pipeline does five things in sequence:

• Runs your PubMed queries
• Normalizes and deduplicates the records
• Has the AI screen every record's title and abstract against the inclusion criteria
• Downloads open-access full text for every record the AI ranks above the cutoff
• Has the AI read each downloaded full text and recommend Include or Exclude

You don't need to do anything during this. The progress page shows live ticks — “48 of 124 records screened” — so you can tell it's working. Most runs complete in five to fifteen minutes depending on volume.

When the pipeline finishes, the page will show a blue button sending you to the next step.

From the breadcrumb at the top of the dashboard, click Abstract review. You'll see a list of records the AI thought were worth your attention, ranked by relevance.

For each record, the AI shows you its recommendation, its confidence, and a short rationale. Read the title and abstract. If you want this study in your review, click Include. If you don't, scroll past it. That's the entire interaction.

The toolbar at the top of the inbox has a Select all AI-includes shortcut if you want to bulk-confirm everything the AI flagged as a clear include — useful when you've worked with the AI long enough to trust its high-confidence calls. You can always un-include individual records afterward.

From the breadcrumb, click Full-text review. This queue contains every record you Included at the abstract stage for which the system was able to retrieve full text.

Each record card has an Open full textlink that opens the parsed plain text in a new tab. The AI's full-text-level recommendation is shown alongside, with a rationale that quotes the article.

Read what you need to read. If the article belongs in your evidence table, click Include. If not, scroll past. Same rule as abstract review.

From the dashboard, click Extract evidence for N records →. This is the only stage that does actively require a click on every record — every Included full-text article needs an appraisal sign-off because the appraisal text is what lands in the 1.10.3 Evidence Table.

The AI has already drafted the appraisal: study design, methods, key findings, comments and limitations, PMTA domains, materiality. Click any row to expand it and review the draft.

Edit any field that needs correction — scientist text always wins over AI text. When the appraisal reads correctly, click Save & next. The row collapses, the next pending row expands, and the page scrolls to it. You don't need to navigate away.

If the AI couldn't read the full text for some reason — you'll see an amber notice and a Retry AI extraction button on that record. Click it; the page polls for the new draft and populates the fields when it lands. No manual refresh needed.

If your review is single-scientist, you can skip this section entirely — adjudication only appears when two scientists have recorded conflicting decisions on the same record.

When it does appear, the dashboard will show a Resolve N conflicts → button. The adjudication page lays out both scientists' decisions side-by-side with their rationales. You make a third call that supersedes the first two.

With every appraisal signed off, the dashboard's primary button becomes Build outputs →. Click it.

The build pipeline runs six stages in sequence: it generates the PMTA narrative draft, catalogs article access, builds the screening audit trail, computes the PRISMA flow counts, finalizes the 1.10.3 Evidence Table, and renders the report PDFs. The progress page ticks per section as the narrative synthesis runs — you can tell live whether it's working or stalled.

Builds typically finish in two to four minutes. When they do, the page sends you straight to the downloads.

The downloads page lists every artifact the build produced. Click any filename to save it locally. The packet you hand off to the regulatory team typically includes:

• narrative_draft.md — the 1.10.2 narrative scaffold, scientist edits go here
• evidence_table_template.csv — the 1.10.3 Evidence Table
• prisma_counts.json / .csv / .md — the PRISMA flow counts for the methods section
• title_abstract_screening.csv and full_text_review.csv — the audit-trail workbooks
• referenced_literature_manifest.csv — every cited record with its access provenance

Every error in the workspace is recoverable. The general rule is: if you see a red error message, look for a blue retry button next to it. The two most common cases:

• A pipeline stage failed. The progress page will show Retry pipeline. Click it; the run picks up cleanly from the start. Your scientist decisions and appraisals are preserved — they live in a different table from the pipeline state.
• A build stage failed. The progress page will show Retry build. Click it; only the build re-runs. None of your screening work is touched.

If a retry fails twice in a row, that's when to flag it. The error message and the “Show full traceback” link below it carry enough context for the engineering team to diagnose without you needing to reproduce.

The fastest path to confidence is doing a single end-to-end run on a small update window — say, the last quarter — and seeing the deliverables land. Once you've seen the workspace produce one packet, the rest of the protocol cycles will feel familiar.

Sign in to start →