Delphi
What the Greeks knew.
Delphi sat on the slopes of Mount Parnassus, considered by the Greeks to be the omphalos — the navel of the world. Its priestess, the Pythia, delivered prophecies in a trance state above a chasm in the earth. Modern geology suggests volcanic gases rose through the rock and induced her altered consciousness. Her utterances were ambiguous on purpose, polished into verse by attending priests, and notoriously open to multiple readings.
Kings and generals consulted Delphi before launching campaigns. City-states consulted Delphi before founding colonies. Philosophers consulted Delphi about questions of meaning and self. The discipline of consulting Delphi was the discipline of preparing to hear an answer you would have to interpret.
Croesus asked whether to attack Persia. The oracle replied that a great empire would fall. He attacked. The empire that fell was his own.
That story is the heart of Delphi. The Pythia's answer was correct. It was also unreadable in the moment, because the question Croesus asked — "will I succeed?" — was answered with a different question's frame — "what will happen?" Delphi did not tell you what to do. Delphi told you what was at stake, and left the reading to you.
This is the oracle of high-stakes interpretation. The one whose answers reward careful reading, weigh competing meanings, and resist easy reduction. When the question is subtle, Delphi is the one you want on the panel.
Replicating Delphi.
Our Delphi is the flagship reasoner on the panel. When an article needs to be evaluated for nuance — tone, voice, the experience signals that don't reduce to a checklist — Delphi reads it carefully, considers multiple interpretations, and renders a verdict that respects the ambiguity in the source.
The implementation uses the most capable language model available to us, configured for deliberate, multi-pass reading. The system prompt instructs the model to do what the Pythia's priests did: read for what is said, read for what is implied, and weigh the difference. Where another oracle might return a single confident answer, Delphi is asked to surface tension. To name what is uncertain. To resist the temptation to simplify.
The trade-off is real. Delphi is slower than the others. Delphi sometimes disagrees with itself across runs — the same way the Pythia's answers were never quite the same twice, even on similar questions. That variance is part of the design. It is also why Delphi is one of five, not the whole panel. A single Delphi reading is one informed interpretation. Five oracles reading together is the verdict.
When you see Delphi's score on your audit, you are seeing what a careful, considered reader thought of your article on this particular reading. Not a measurement. An interpretation. With dissent from the other four oracles surfaced when it appears.