You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
gbrain's headline pitch is "search gives you raw pages; gbrain gives you the answer." Today kb.context returns a ranked list of approved claims with citations — useful, but the calling agent still has to synthesise. The natural vouch shape, distinct from gbrain's, is review-gated synthesis: answer only from approved claims, surface explicit gaps ("the KB doesn't know X"), and never hallucinate around a missing citation.
This is the single biggest functional differentiator vouch lacks. Done right, it turns the review gate from a constraint into a feature: every word in the synthesis traces to a claim a human approved.
Suggested shape
vouch synthesize "what's our auth model?" [--depth 3] [--max-chars 4000]
# CLI; same surface available as kb.synthesize over MCP + JSONL
New src/vouch/synthesize.py builds an answer by walking the existing kb.context ranking, then composing a paragraph per relevant cluster with inline [<claim_id>] citations.
Refuses to synthesise across uncited spans; every sentence carries at least one citation.
A trailing gaps: block lists topics the query implies but no approved claim covers (gaps: ["mfa", "session-rotation"]).
Configurable backend: deterministic (template-driven) by default; --llm opts into an LLM-mediated rewrite that still must cite each sentence — the deterministic path is the audit-friendly default.
_meta.synthesis_confidence field on the response: high (every sentence cited from a stable claim), medium (some working-status claims), low (contested claims present).
Acceptance
A KB with 3 approved claims about auth produces a 1-paragraph synthesis citing all 3 by id.
A query the KB doesn't cover produces an empty body + a populated gaps: block.
Synthesis never includes text not traceable to an approved claim (a fuzz test asserts this).
What you're trying to do
gbrain's headline pitch is "search gives you raw pages; gbrain gives you the answer." Today
kb.contextreturns a ranked list of approved claims with citations — useful, but the calling agent still has to synthesise. The natural vouch shape, distinct from gbrain's, is review-gated synthesis: answer only from approved claims, surface explicit gaps ("the KB doesn't know X"), and never hallucinate around a missing citation.This is the single biggest functional differentiator vouch lacks. Done right, it turns the review gate from a constraint into a feature: every word in the synthesis traces to a claim a human approved.
Suggested shape
src/vouch/synthesize.pybuilds an answer by walking the existingkb.contextranking, then composing a paragraph per relevant cluster with inline[<claim_id>]citations.gaps:block lists topics the query implies but no approved claim covers (gaps: ["mfa", "session-rotation"]).--llmopts into an LLM-mediated rewrite that still must cite each sentence — the deterministic path is the audit-friendly default._meta.synthesis_confidencefield on the response:high(every sentence cited from a stable claim),medium(some working-status claims),low(contested claims present).Acceptance
gaps:block.kb.synthesizeis reachable over MCP stdio, JSONL, HTTP (chore(deps): bump actions/setup-python from 5 to 6 #1), andkb.capabilitieslists it.Out of scope