Part 2 of 2. Part 1 was about the questions a model simply can't answer. This one is about the larger, sneakier set, where the model hands you a perfectly plausible answer, and the problem is everything it quietly left out.
In Part 1, the failure was obvious: ask a model for the deep roster of a field and it runs dry or invents. Most real questions are not like that. Ask a model to "build me a panel" or "give me ten experts to call" and it produces a confident, reasonable-looking list. That is the dangerous case, because the list looks complete, and you have no way to see what is missing from it.
The fix is not to replace the model. It is to let the model draft freely, then run its answer through the Authority Index as a verification-and-completion pass: confirm the names are real, and surface the people the model could not see it had omitted.
The loop
- Claude drafts a first-pass answer from its own knowledge.
- Authorix
matchchecks the drafted names: real authorities, and in which lane. - Authorix
indexretrieves the authorities on the topic the draft missed. - Claude revises into a final answer that is verified and complete.
A worked example: a panel on the future of nuclear energy
We asked a frontier model, unaided, to assemble an eight-person panel. It produced a real, credible list: Jacopo Buongiorno, Per Peterson, Kathryn Huff, Rachel Slaybaugh, Ashley Finan, Leslie Dewan, Isabelle Boemeke, Mark Nelson. Nothing invented. But look at the shape of it: every name is US-based, every name is pro-nuclear, and the mix skews to academics and advocates.
Then we ran the draft through Authorix.
names: 8 draftedtopic: "future of nuclear energy"| Name | Affiliation | Focus |
|---|---|---|
| Anil Kakodkar | Former Chairman, Atomic Energy Commission of India | Thorium programme |
| Valérie Faudon | Executive Director, French Nuclear Society (SFEN) | Nuclear policy |
| Jacob DeWitte | Co-founder & CEO, Oklo | Advanced fission |
| Ville Tulkki | VTT, Finland | Small modular reactors |
| Edwin Lyman | Union of Concerned Scientists | Reactor safety (credible critic) |
What we measured
Two things, on this one query.
The draft was real but narrow. Every drafted name verified as a real person, zero fabrications. But the index's bios revealed the composition the model could not: around 5 domain authorities and around 3 advocates and communicators, all from one country and one stance.
| For this panel | Claude's draft | After the Authorix cross-check |
|---|---|---|
| Drafted names verified real | 8 / 8 | 8 / 8 |
| Countries represented | 1 (US) | many |
| Stances represented | pro-nuclear only | pro and credible critics |
| Industry / startup operators | ~1 | many |
One query, one run. Counts of added experts are approximate tallies.
Verification plus right-sizing. The cross-check did not catch a fake, the draft had none. Its value was completion: turning a real-but-lopsided list into a balanced one, with the bios to know which seat each person fills.
What this does, and does not, claim
As in Part 1, the honest line.
The index returns real people, and the ones the model missed. It does not claim a perfect ordering of who matters most, or that every result is a tight topical match.
What this does and does not claim
What it reliably fixes is the blind spot: the geographies, sectors, and dissenting voices a model's confident first draft silently omits. What we are not claiming: a perfect ordering of who matters most, or that every result is a tight topical match. Ranking and relevance precision are work in progress.
That is the whole point of the loop. The model is good at producing a fluent first answer; it is bad at knowing what it left out. The index is the opposite. Together they produce an answer that is both well-reasoned and complete, which neither does alone.
The two posts together
Part 1: the questions a model cannot answer, where the index is the only source. Part 2: the answers a model gives confidently, where the index catches the blind spots. Same underlying fix in both, ground the model in verified, real-world authority data, pointed at the two different ways a model falls short.
Draft, then cross-check