How it works
Briefly uses a two-pass AI pipeline plus risk scoring. Every claim in the brief and in the step-by-step view is tied to verbatim evidence from the document.
PASS 1
Extractor
The document is cleaned and optionally chunked. The Extractor pass runs per chunk and outputs structured facts: category, claim, and evidence quotes (with offsets when possible). No summaries—facts only. If something isn't in the document, we don't invent it.
MERGE
Merge engine
Chunk outputs are merged: deduplicated by claim/category, with conflicts resolved by choosing the fact with the strongest evidence (verbatim quotes, optional offsets). Document type (ToS, privacy, contract, etc.) is chosen by confidence across chunks.
RISK
Risk scoring
A rule-based risk engine scores patterns in the merged facts: arbitration, auto-renewal, broad liability, data selling/sharing, unilateral changes, etc. Combined with writer-reported severity, we output a single Low / Medium / High score.
PASS 2
Writer
The Writer receives only the merged extraction. It produces the 100-word brief, 3–8 key flags, and step-by-step sections. It is constrained to not add any claim not supported by the extracted facts. Every flag and every section item must include at least one evidence quote.
ENFORCEMENT
Exact 100-word brief
The brief is enforced to be exactly 100 words. If the Writer outputs more, we apply deterministic compression (remove filler phrases, then trim to 100 words). If needed, a constrained "compress-only" LLM step can run without introducing new information. The result is validated against the extraction.
Flow diagram
Document → Clean & Chunk → Extractor (per chunk) → Merge → Risk Engine → Writer → 100-word enforcement → Brief + Flags + Sections (all with evidence).