A citation is a starting point, not a verdict
Turn documents into a reviewable brief.
Bring a question and your sources. Take a prompt to your model, bring back its answer, then inspect each claim beside the exact words it cites. Export only the claims you approve—with unresolved questions and an honest review appendix.
Local-only · no account · no model calls · no content upload · no autosave
Return after a source changes: keep a review snapshot, update a versioned document and inspect which earlier claims need attention.
See a real-source Python migration brief: two pinned documentation versions, an actual local runtime probe, and a rejected claim that still passes quote matching. These are agent-authored editorial judgments, not a model benchmark.
The workspace requires JavaScript. The worked example and downloads do not.
1. Set the question and collect sources
Paste text you have permission to use. The tool does not fetch URLs, parse PDFs, browse the web or establish provenance. Keep version/date/page information in each source’s locator. Compare historical versions as historical versions, not automatically current facts.
Load Python documentation example
Clear this workspace
Open evidence-brief work JSON (UTF-8, up to 2 MB)
Opening asks before replacing work. No saved approval is restored. This format is separate from Test Bench projects and model answers.
Brief title
Question and intended scope
Be specific about audience, versions and what evidence would answer the question.
Add source
Up to 12 sources, 60,000 characters each. IDs stay attached to their source when another source is removed; they are not silently renumbered. Source text and metadata edits clear every current claim decision. Whitespace is significant in quote matching; native textarea editing may normalize pasted line endings.
Download source-grounded prompt
Copy prompt
Inspect the exact prompt before sharing
Refresh prompt preview
The prompt includes your question, every source title/locator and full source text. It excludes the brief title, rights fields, notes and current answer. Your external model service has its own privacy and costs. The tool does not record or authenticate a model run.
2. Bring back an answer, check its citations
Raw JSON answer (up to 160,000 characters)
Answer format and check limits
{
"claims": [{
"id": "C1",
"text": "One specific claim supported by these sources.",
"citations": [{"source": "S1", "quote": "Exact contiguous text from source S1."}]
}],
"questions": ["What evidence is still needed?"]
}
Up to 40 claims, 8 citations per claim and 30 questions. Missing source IDs or nonmatching quotes block that claim, not the whole answer. A claim with no citations cannot be approved. A malformed response blocks preview. Duplicate JSON keys follow native parsing (last value wins); inspect the parsed preview. Matching does not judge entailment, completeness, source quality, conflicting versions or truth.
Preview claims and citations
Nothing saved or sent. Add sources, then preview an answer.
3. Keep the reasoning, share a scoped brief
Review notes, rejection reasons and next actions
Notes appear in both work files and briefs. Editing notes clears current approvals so the exported interpretation is reviewed again. Use them to explain disagreements; decisions alone are not a review record.
Download reviewed HTML brief
Download Markdown brief
Save work JSON
HTML is a read-only, script-free offline document; Markdown is editable reading material. Both contain approved claims with quotes and source locators, unresolved questions, excluded claim text/status, your notes and all source metadata and rights fields. They do not include full source bodies or the raw answer. Inspect the review appendix and expandable rights before sharing. They are not re-importable work files or proof of verified facts.
Work JSON retains all current source bodies, metadata, rights, the raw answer even if invalid, question, title and notes. It does not restore current claim decisions. Optional captured snapshots below retain historical evidence and decisions separately; reopening always requires a fresh preview and individual approval. Ordinary brief/prompt exports use only current work, not snapshots. Keep downloaded reading briefs separately if you need their original bytes. Files are unsigned, unencrypted and unredacted; no source authenticity or execution identity is verified. Reloading loses this tab’s work unless you download it. Clear removes tab contents, not previously shared/downloaded files or clipboard text.
4. Return after a source changes
Capture the evidence behind a review before editing documents. A snapshot keeps the full current question, source bodies/metadata/rights, raw answer, notes and every claim’s decision—including holds and rejections. It does not capture earlier snapshots, the exact prompt sent to a model, model execution or authenticated identity. Snapshot dates come from your device and are unverified.
Snapshot label (up to 120 characters)
Capture current review snapshot
Preview and make at least one explicit approve/reject decision first. Capture stays in this tab until you Save work JSON. Up to five snapshots within the same 2 MB file budget; no automatic trimming. Work with snapshots uses version2; older editors reject it. Work without snapshots stays version1.
Old private text survives edits and source deletion. Snapshots retain complete original sources, raw answers, rights and notes. They are included in work JSON, not current prompts/briefs. Delete a snapshot explicitly to remove its retained data from this working copy; older downloads still contain it. All current approvals remain separate and are never restored when opening work.
Retained review snapshot
Compare with current evidence
Generate historical HTML brief
Download snapshot evidence only
Delete selected snapshot
Historical HTML uses that snapshot’s decisions, not current approval; it needs at least one prior approved claim. It is generated now, not guaranteed to reproduce an old download byte-for-byte. Snapshot-evidence JSON contains that one earlier working copy without decisions or other snapshots; save current work before opening it. A comparison matches exact IDs only, never guesses renamed identities. Source changes can matter even when old quotes still match. Exact restoration can match again, so this is not a permanent audit log.
How to review a claim
Check the source identity and version. A perfect quote from the wrong version can still mislead.
Read the cited words and their surrounding source. Multiple exact occurrences are disclosed; the line range is the first match in your pasted text.
Ask whether the quotes support every part of the claim. A source saying “does not write” cannot support “can write,” even though its quote matches.
Reject unsupported conclusions. Keep missing evidence and conflicts as questions rather than making the brief appear complete.
Approve claims individually, then read the exported brief as a recipient. Recheck included private text and attribution before sharing.
Source Brief is a separate task workspace, not a quality score, citation crawler, formal research system or Test Bench import. Its limits and review policy are provisional product choices. Privacy details.