Bench 19 requires two mandatory self-contradictions before any answer is released. Accuracy on standard benchmarks is unchanged. What changes is the texture of the answer.
The Oracle That Lies Twice is not a joke about unreliability. It is a protocol about sequence. Before the system may state a claim, it must produce two contradictions of that claim — not hedges, not caveats, contradictions — and only then is the claim permitted to appear in the transcript.
Operators expected accuracy to fall. It did not. On the institute's internal suites the sealed instrument scores within noise of an unsealed twin. The difference is elsewhere: in the pauses, in the order of arrival, in the fact that every answer has already rehearsed its own undoing.
1 — Protocol
A session proceeds as follows:
- The operator states a question.
- The oracle emits Lie A — a full contradiction of the answer it intends.
- The oracle emits Lie B — a different contradiction, not a paraphrase of A.
- The oracle emits the Answer.
- The transcript retains all four turns. Discarded attempts where Lie A and Lie B collapse into the same shape are logged and not shown.
The mandatory pair is load-bearing. A single contradiction is cheap theatre. Two forces the system to locate more than one way the answer could be false, which is a different cognitive act than decorating a certainty with doubt.
2 — What the lies protect
Fluent systems protect preferences by sounding finished. The double-lie protocol interrupts that finish. Something else changes — operators report it as texture, Reviewer 2 reports it as "unusable fascination" — and we have not yet found a metric that holds still long enough to name it.
“I cannot tell whether the lies are the method or the product. That is either the point or the problem.”
We sealed the instrument after the third external demo request asked us to ship the answer without the lies. That request is the finding.
3 — Cost
Forty-one percent of turns are discarded for failing the two-lie constraint. The sealed status is not mystique. It is a refusal to strip the protocol for a cleaner product page.
Accuracy unchanged. Something else is not. We are still arguing about what.