Coming Soon · Free
Make your knowledge AI-searchable — by proving a real question grounds or refuses, not by tidying folders.
A working session for the operator whose AI agent gives confident, wrong answers about their own documents. You inventory the corpus you actually have, design a taxonomy an agent retrieves against — entities, relationships, retrieval keys, chunk shape — and then run the grounding test: real questions answered FROM the structure, or correctly refused when the answer isn't there. You leave with a Taxonomy + KB Shape artifact and a grounding verdict. The thesis: a taxonomy you haven't tested is a hypothesis, not a result. No grounding test, no AI-ready verdict.
The arc
Name the one knowledge domain in scope and show ONE question the AI gets wrong about it today. Separate the retrieval gap (can't find) from the grounding gap (found wrong / invented) — it calibrates how hard everything downstream probes.
Surface the real document sources — location, owner, format, volume, freshness — fast, not exhaustive. Then the load-bearing move: bet-your-job source-of-truth resolution, plus the scope-exclusion edge the adversarial test will land outside of.
Build the structure an agent retrieves against — entities, relationships, retrieval keys, chunk shape — harvested BACKWARD from real questions and their answer-shaped passages, never drawn top-down. Each entity resolves to one authoritative source; each key reverse-engineered from a real query.
Also from Framework Friday
What you'll walk away with
Score your workflows against five operator-tier criteria, map your top priority end-to-end, and produce a spec with governance that a builder could pick up tomorrow. Yours to keep, refine, and put to work — whichever way you choose to run it.
THE WORKSHOP. Run 3+ real questions against the structure — a factual lookup, a cross-entity question, and the non-optional adversarial one the corpus can't answer — and verdict each ANSWERED / CORRECTLY-REFUSED / HALLUCINATED on structural evidence, not vibe. The only phase that can stamp the corpus AI-ready.
Synthesize the four-section Taxonomy + KB Shape artifact, run the four binary-screen stamps (A-D) out loud, deliver the prep-action list ordered by hallucination RISK (confident-wrong before blanks), declare the governance exemption inline, and frame the FridayOS ds-import bridge as completion, not handoff.