Built 2026
Knowledge-base rebuild
When the reasoning survives but the evidence doesn’t
Outcome
Rebuilt a knowledge base that looked complete and was effectively useless — good structure and narrative, but most of the document links behind it were dead. Every claim now traces to a source document that can actually be produced; what couldn’t be sourced is flagged rather than quietly left standing.
How it was built
The original had the reasoning but not the evidence — the arguments survived, the links to the underlying documents didn’t. Rebuilt by re-deriving each claim from a document that can be produced, flagging what can’t be sourced, and never editing the original. A verification pass, not a trust pass.
What it demonstrates
The clearest example of the grounding problem — a system confidently useless because it summarized its own context instead of being anchored to source. The fix wasn’t a better tool; it was reconnecting the knowledge layer to real material, claim by claim. This is the single best illustration of what I actually fix, and the problem is completely generic: any team whose internal copilot or wiki is “almost right but not quite” has this.
Call me when
You have a knowledge base, wiki, or internal AI assistant that sounds authoritative but can’t be trusted — where the answers look right but nothing underneath them is verified.
Contact
If this sounds like a problem you have, the best place to start is a short conversation about what’s actually stuck.
Email me