The Briefing Engine generates auditable decision templates with structured evidence, explicit conflicts, and clear boundaries.
The engine splits regulatory documents into overlapping text chunks – e.g., each provision of the UK Code as its own unit. Context is preserved.
Each chunk is turned into a vector (384 numbers). The model "all-MiniLM-L6-v2" captures meaning – not just keywords.
Your question is mapped to the same vector space. The engine finds the 3 most similar chunks – by meaning, not keywords.
From the chunks we build an evidence pack: original text, source, evidence class (Gold / Regulatory / Web), confidence, date.
The evidence pack is mapped to the seven mandatory fields. Only if all are fulfilled does a briefing appear.
If the top chunks don't provide sufficient evidence (question too specific, no source), the engine does not hallucinate. It returns HTTP 422 – ABORT with explanation.
Every briefing contains a full audit trail: which source, which chunk, which model, which timestamp. Every claim is verifiable.
🧪 Real example: Question about "Independence requirements for audit committee chair under UK Code 2024"
The engine searches 25 chunks of UK provisions. The most similar chunk is D.4 (Audit Committee of independent non‑executive directors). The briefing quotes the original text, assigns evidence class GOLD FIXTURE, source FRC. Uncertainty is rated MEDIUM because the interpretation of "independence" can be company‑specific.
I'm looking for 3 pilot customers who will actually use the system. No hidden costs, no lock‑in.
Ask a governance question and see the 7 mandatory fields fill in real time.
Try keywords like "NIS2", "CSRD", "SOX", "risk management"