Cumae
What the Greeks knew.
The Cumaean Sibyl lived in a cavern carved into a volcanic hillside near what is now Naples. Her cave was long and narrow, cut into the rock by hands no one remembered, with a hundred openings to the outside. When she spoke under Apollo's inspiration, her voice echoed through all hundred openings at once, as if the hillside itself were prophesying.
She did not deliver her prophecies aloud. She wrote them. On individual oak leaves. Each prophecy on its own leaf, arranged carefully across the cavern floor in the order she had set them down.
The catch was famous, and it was the reason consulting Cumae was both prized and risky. The cave had heavy doors, and when petitioners opened those doors to enter, drafts of wind would sometimes catch the leaves. The prophecy was still there — every leaf still in the cave, every word still written. But the order was gone. The truth had scattered.
Reading the Sibyl required assembling the fragments back into a coherent answer. The truth was real. It was also broken into pieces by the act of approaching it.
The Sibyl herself would not reassemble the leaves. She would not even tell you which pieces were missing. Petitioners gathered what they could, sat with the fragments, and worked out the prophecy from what survived. Some left convinced they had read the answer. Some left knowing they had only read part of it. Most never knew which they had done.
Centuries later, the Romans collected texts attributed to her — the Sibylline Books — and consulted them in moments of national crisis. They were treated as so important that they were guarded by a dedicated college of priests, and consulted only by order of the Senate. The fragments, gathered and preserved, became one of the most important reference works in the ancient world.
Replicating Cumae.
Our Cumae is the cross-referencer on the panel. Where the other oracles read your article in isolation, Cumae reads it against the wider world — checking claims against external sources, verifying citations against original documents, finding what other authorities have said about the same questions.
The implementation uses a search-augmented model: a language model that can query the open web before answering. When your article makes a factual claim, Cumae consults the leaves — the published, indexed, retrievable record of what is known — and reports whether the claim survives that consultation. The structural similarity to the Sibyl's method is almost exact. The fragments exist; they are scattered; Cumae's job is to gather what is relevant and assemble a coherent reading.
This oracle is structurally different from the others in a way worth naming. The other four are parametric — they reason from what they were trained on, which is fixed at the moment of training. Cumae is retrieval-augmented — she goes outside her own memory to consult sources that exist in the live world. That asymmetric capability is the reason she is on the panel. A panel of five parametric oracles would share the same blind spots about anything new, anything specialized, anything the training data did not capture well. Cumae covers what they cannot.
The Sibyl's leaves were sometimes incomplete after scattering. Cumae's web search is sometimes incomplete too. There are claims that no source on the open web can verify. When Cumae reports that a claim is unverified rather than verified or contradicted, that report is itself useful. The leaves did not contain the answer. Sometimes that is the prophecy.
When you see Cumae's verdict on your audit, you are seeing what the published world says about the claims in your article. Not what one model believes. What the leaves — gathered, fragmentary, real — can confirm or contradict.