Guide 3 — The Oracle
The Oracle and Ask Interface
Citation-grounded question-answering, with deliberate abstention
This documentation describes Theseus Codex's infrastructure and methodology. It does not expose private firm materials, uploaded source documents, or unreleased internal records.
- For
- Readers who want to understand what happens when the firm answers a question.
- Summary
- The Oracle is the firm's question-answering surface. It does not run a free chat against a foundation model; it answers from the firm's corpus and refuses, or warns, when the corpus is too thin. Every quoted span must appear verbatim in the cited source — if it does not, the answer is rejected before it reaches the reader.
How an answer is composed
The path under the hood is short and conservative.
- Encode the question into a fingerprint.
- Retrieve the top-ranked principles and conclusions nearest that fingerprint, filtered by a minimum conviction threshold.
- Pull the supporting claims for each retrieved principle.
- Hand the principles (as axioms) and claims (as evidence) to a language model with a strict contract: produce an answer, cite which principle grounds each part, note caveats.
- Validate that the answer does not contradict the retrieved principles with a coherence check.
- Generate an adversarial counter-position to the leaned-on principles and include its strongest objection as a caveat.
- Run a final consistency check before returning the answer.
The verbatim-citation rule
The single load-bearing rule of every public-facing surface is the verbatim citation contract: any quoted span the answer leans on must appear word-for-word in the cited source. The validator checks this on every answer. When a quoted span does not match, the answer is not retried with a fix-it prompt — it is rejected, and the Oracle abstains.
The same rule echoes through the stack. Currents abstains rather than publish an opinion whose quote is not in any source. Forecasts abstains rather than ship a prediction whose reasoning cites a span that is not there.
Confidence bands
Each answer carries a band based on the strength and breadth of the citation chain.
- high — supported by multiple independent citations, including at least one principle or conclusion.
- medium — supported by at least one citation that directly addresses the question.
- low — citations exist but only weakly support the answer; treat as a starting point.
- abstain — the corpus is too thin or off-topic to answer. The Oracle returns no answer text and explains what material it would have needed.
Hallucination warning, and what to do with low confidence
When post-hoc verification finds the answer contains a span the cited sources do not unambiguously support, the system surfaces a hallucination badge. The badge does not mean the answer is wrong; it means the support is weaker than the prose makes it sound. The recommended response is to read the cited sources, decide whether the unsupported span is a paraphrase that survives, and narrow the question if it does not.
Low confidence does not mean probably wrong. It means the corpus will not let the Oracle commit to a sharper answer. Standard responses are to add corpus on the topic, narrow the question, or accept the abstention.
Inverse inference
A specialty mode runs the inverse question: given that an event happened, what does it imply about (or refute in) the firm's existing positions? The result is a list of supporting implications, refuted implications with severity, a transparency list of irrelevant principles, and a blindspot report of entities, mechanisms, or adjacent topics the corpus is missing in light of the event.