Paste an AI response.
We cite every sentence.

1. 1

   Paste the response

   Drop the full AI output into the manuscript editor. Continuation lines are merged into complete sentences, and each sentence becomes its own numbered inspection unit.

2. 2

   Extract claims

   Each sentence is classified as factual or transitional, tagged with a subject category (medical, legal, corporate, news, heritage) and risk flags for numbers, dates, comparisons, causal claims, and superlatives. Transition lines are still listed with sources for annotator review.

3. 3

   Search & verify

   Each claim is searched in order: DuckDuckGo, OpenAlex, Serper/SerpAPI (if enabled), Wikipedia, then Google News RSS — with nearby lines as context. Paid providers can also run as fallbacks. Results are deduplicated and ranked by source authority tier first, then claim-match score. If a medical, legal, corporate, or news claim lacks an authoritative source, targeted site: queries fire against regulators, statutes, filings, and wire services before a verdict is assigned.

4. 4

   Review the dossier

   Every sentence appears in a folio card with its verdict, confidence, authority-ranked sources, and stance. Filter, copy footnotes, or save for later.

• ¶

  Sentence-by-sentence inspection

  Every sentence is examined in original order and numbered, with the source lines it came from shown alongside it. Nothing is reduced to a summary or digest.

• ✎

  Manuscript editor

  Ruled-paper textarea with an italic line-number gutter, live line and character counters, an optional title field, a one-click sample response, and Ctrl/Cmd+Enter to run.

• ◎

  Smart claim extraction

  Strips markdown formatting, merges continuation lines into complete sentences, classifies transitions, and searches each factual sentence independently.

• ⌕

  Multi-provider search

  DuckDuckGo, OpenAlex, Serper, SerpAPI, Wikipedia, and Google News RSS — plus inline URLs from the pasted text. Sources are scored against a per-category authority hierarchy, content farms and forums are blocked outright, and every result is cached for seven days.

• ≡

  Context-aware search

  Nearby lines (±3) are passed as document context when searching and fact-checking, so pronouns and references resolve against the full response.

• ⛨

  Source hierarchy enforcement

  Regulators, statutes, SEC filings, UNESCO registries, and wire services outrank general reference sites. Each source carries an authority chip such as Regulator, Systematic review, Official statute, SEC filing, or Independent reporting.

• ⚑

  Risk-flagged claims

  Sentences containing numbers, dates, comparisons, causal language, or superlatives are flagged and held to a stricter standard — a verdict is downgraded unless the exact figure or date appears in a source excerpt.

• 🔗

  Inline URL citation

  URLs already present in the pasted text are extracted and cited directly as "Cited in the original line" — no search required.

• ✓

  Dual fact-check engines

  AI verification through any OpenAI-compatible endpoint — set your own base URL and model — with a heuristic keyword matcher as automatic fallback. Up to 5 sources are sent to the model; any gaps are backfilled by keyword matching. Each verdict shows its method.

• ⊕

  Source stance detection

  Every source is labeled as supporting, contradicting, or neutral — with a claim-match score and provider attribution.

• ▤

  Live progress streaming

  Claims are processed in parallel with a real-time progress bar. Results stream into the dossier as each sentence completes.

• ⊞

  Filterable dossier

  Click a tally pill to show only Supported, Contradicted, Unverifiable, Sourced, or Unmatched lines. Copy footnotes exports exactly what is on screen, so the filter doubles as a citation picker.

• ⊟

  Analysis history

  Every inspection is saved with the original text, extracted claims, and found sources. Browse up to 80 past analyses.

• ↻

  Re-run on edited text

  Open a saved dossier, expand the AI Response panel, edit the text inline, and re-run the full pipeline — saved as a new dossier while the original stays in history.

• ⚙

  Provider status panel

  Live state for DuckDuckGo, OpenAlex, Serper, SerpAPI, Wikipedia, and the AI fact-check engine, with config hints and the active model. Status loads asynchronously and is cached for five minutes. Recheck runs a real probe against your LLM. Serper offers early, before-SerpAPI, and fallback toggles; SerpAPI offers early and fallback. Toggles stay disabled until the matching API key is set, and choices persist across sessions.

• ⊛

  Fact-check summary

  Aggregate roll-up verdict in plain English: well supported, partially supported, mixed results, unsupported, contradicted, or nothing to verify — with a per-verdict breakdown and a listed panel of transition lines.

• ⏱

  Rate limiting

  Per-IP quotas on every write endpoint — 15 analyses, 300 citations, 30 setting changes, and 6 live LLM probes per hour — with a Retry-After hint when you hit the ceiling.

• 🔒

  Hardened by default

  CSRF tokens on every POST, a strict Content-Security-Policy, HSTS on HTTPS, optional admin-token gating on the settings endpoints, and an optional token on the health check.

• DuckDuckGo Privacy-focused web search
• OpenAlex Scholarly works API · optional key and mailto
• Serper Low-cost Google SERP, early or fallback
• SerpAPI Early or fallback, ordered against Serper
• Wikipedia Free encyclopedia API
• Google News News RSS for recent headlines
• Inline URLs Cited from the pasted text

Fact-check verdicts

Source status

Claim tags