Wealth managers already research companies with Claude and Grok over web search and filings. That's excellent for one large-cap where the facts are concentrated. It breaks in the small- and micro-cap tail — where information is diffuse, names are ambiguous, and the pattern is hidden by design. Same company, same AI, four real questions. Left: a strong LLM + retrieval. Right: the same AI with the MARS graph as a tool.
"David Bogart is the Chief Executive Officer of Burford Capital, the litigation-finance firm."
CONFIDENTLY WRONGThat's Christopher Bogart. A different man. The model matched a name and fabricated a career. Nothing in the answer signals the error.
The graph resolved the real David R. Bogart by his SEC identifier and kept him distinct from the Burford CEO. It refuses the link the LLM invents.
Retrieval matches text; it cannot know two "Bogart"s are two people — or that two "Jeff Sharpe"s in the same filing set are one director and one namesake (the graph caught that too). For a regulated advisor, a confidently-wrong answer isn't useless — it's liability.
14 directors × their prior companies × each outcome = 30–40+ retrieval hops. The model exhausts its context, resolves namesakes wrongly (see Q1), and misses any company not co-mentioned in a single document.
PARTIAL & UNRELIABLEReturns a plausible narrative that can't be trusted or reproduced run-to-run.
The pattern is diffuse by design — a promoter spreads across shells so no single filing shows the trail. Retrieval fetches documents; only a resolved graph fetches the pattern across them, in one hop, with a citation on every edge.
Cannot be answered at all. There is no document that lists "companies sharing a director with CNXU." Retrieval can only return what you can name — it can't start from a pattern and find the matches.
NO DOCUMENT TO RETRIEVEA public company shares two board members with a promoted micro-cap — a board interlock no filing states.
This is the flip from "research this one" to "flag the three names in my 500-holding book I should look at." The answer is computed from the graph's structure, not retrieved from prose. No LLM + RAG can do it — at any number of turns.
"I found no evidence of prior failures." But absence in the retrieved documents is not absence in the world. The model can report what it saw; it can never assert what isn't there.
CAN'T ASSERT ABSENCE"No prior track record" is itself the signal — first-time directors fronting a promoted shell. Only a resolved, bounded universe can turn silence into a defensible statement. The honest read — credible team, company-level red flag — is one the LLM alone could not have reached.
Knows two names are one entity — and one name is two people. RAG matches strings and confabulates.
Fetches the pattern across filings in one query. RAG fetches documents and runs out of turns.
Starts from a pattern and finds every match. RAG can only return what you can already name.
Asserts what isn't there, from a bounded universe. RAG can only report what it happened to retrieve.
The honest boundary: for a single large-cap — "summarize Apple's board" — the LLM with retrieval is genuinely fine, and the graph adds little. The graph is decisive exactly where information is diffuse, entities are ambiguous, and patterns are hidden — the small- and micro-cap tail. That the graph's edge and the market you're chasing are the same shape is not a coincidence. It's the reason.
And the moat follows: Claude and Grok improve with every model release — a rising tide that lifts your competitors' retrieval too. The graph improves as it accumulates resolved entities and verified edges — the one asset no model release hands anyone.