From question to accountable decision record.

Capability is converging. Confidence is not calibrated.

Frontier models are clustering on the benchmarks that used to separate them, and the gap between first and fourth place keeps shrinking with each release cycle. What has not improved at the same rate is calibration. A model's fluency is uncorrelated with its correctness, so a confident wrong answer and a confident right answer are rendered identically. The user is handed the one thing the system cannot supply: the judgment to tell them apart.

A single answer carries no error signal.

One model returns one distribution over one framing. Everything it did not consider, could not check, or quietly assumed is absent from the output — not flagged as missing, simply not there. There is no residual, no disagreement term, nothing to inspect. Retrieval and longer reasoning traces raise the average quality of that answer without producing the one thing missing from it: an independent estimate of when to distrust it.

Independent models fail independently. That is the exploitable structure.

Different pretraining corpora, different post-training, different refusal surfaces, different failure modes. Models trained separately are far more likely to agree when an answer is well-supported than when it is fabricated, because there are many more ways to be wrong than to be right. Correlated failure across independent systems is rarer than correlated success. Disagreement is therefore not noise to be averaged away — it is the measurable signal a single model structurally cannot emit.

A vote throws the signal away. A record keeps it.

Ensembles and majority voting collapse disagreement back into a single number and discard the reason it existed. R.U.D.E. does the opposite: it extracts testable claims, has models cross-examine each other under enforced criteria, and classifies what actually went wrong — fabrication, unsupported claim, weak sourcing, contradiction, incomplete reasoning, honest refusal, poor calibration. Named evidence that cannot be checked earns zero specificity credit, so a confident citation cannot buy a higher rank. The output is an auditable decision record, not a score.

Verification corrects the record without rewriting history.

For explicit current or primary-source requests, every selected model receives the same retrieved source packet before answering. Independent web verification remains a separate user action after the Council ruling. A claim can be supported, contradicted, mixed, or insufficient — and never labeled supported without a valid linked source. When a detail is wrong but the conclusion still holds, the original claim, correction, evidence, and effect on decision confidence stay attached to the record.

The loop compounds. That is the moat.

Everything learned — verified claims, unresolved claims, failed and partially survived challenges, corrections, remaining landmines, and the closest alternative's strongest contribution — is assembled into the strongest next question. The next run does not start over; it starts from the frontier of what survived. Each cycle raises the floor rather than resampling the same distribution.