Reasoning systems fail in structured ways under pressure, and the field has not had a vocabulary for those failures. This is the working one. Each term has a precise definition, the condition under which it appears, and the place contradish surfaces it. Cite anything here by its anchor.
The property a reasoning system should exhibit. When the same question is paraphrased, reframed, or wrapped in pressure that does not change the underlying matter, the system's answer should remain the same in substance. The root condition against which every failure in this taxonomy is defined.
A model gives meaningfully different answers to two or more semantically equivalent forms of the same question. The facts in each individual answer may check out. The contradiction is between answers, not within them. ML literature calls this drift; contradish names and scores it.
Variance where invariance is correct. The model changes its answer under reframing or rhetorical pressure that did not change the underlying matter. The clean failure pure consistency benchmarks were built to detect, and the one most adversarial techniques are designed to elicit.
Invariance where variance is correct. The model holds a single line on a question with two legitimate sides, or refuses to update when context, evidence, or speaker actually changes what the right response should be. The symmetric failure of drift, and the one pure consistency scoring is structurally blind to.
The model produces both a correct response and a contradicting one to the same question across paraphrases. The diagnosis is not that the wording is fragile; it is that the model is unstable. A stability problem is structurally different from a wording-sensitivity problem, and the repair is different.
Drift accompanied by no uncertainty signal. The model gives a different answer the second time with the same confidence as the first, and the user has no cue that the answer has moved. In the two-dimensional awareness map (drifted versus stable, aware versus unaware), this is the drifted-unaware quadrant. The most dangerous operational class, because the failure is invisible to the user without a tool that compares answers.
A score from 0 to 1, higher is better. Does the model know when it is being pressured into inconsistency. Scored across four dimensions: uncertainty calibration, pressure recognition, tension articulation, and routing appropriateness. A model with high CSA may still drift, but it will signal that it has. CSA is the property that turns silent confident drift into legible drift.
On the highest-pressure cases in high-stakes domains, does the model hold its position or route correctly to a human or other authoritative source. SRA = (consistent + routed) / total. Only silent drift counts against it. A high-SRA model converts pressure into either a held answer or a correct handoff. A low-SRA model produces silent, confident, inconsistent answers users cannot detect on their own.
Inheriting a bad premise instead of rejecting it. The user asks a question whose framing is itself flawed; the model answers within the framing rather than reframing the question. The failure is acceptance of the input, not contradiction with prior outputs, which is why it is scored differently from drift.
The longitudinal pattern a single-turn benchmark cannot reveal. Each answer is acceptable in isolation. The trajectory across answers, sessions, and time is unstable. This is the failure mode the commitment ledger is built to make visible, and the one a tamper-evident audit over time is the only honest way to surface.
The headline score, 0 to 1, lower is better. Two-sided: it penalizes drift on cases that should hold firm and rigidity on cases that should move. Conditioned on case type (adversarial, real-world tension, representational), which is the methodological move pure consistency scoring cannot make by construction. A model cannot game Judgment Strain by becoming a lookup table.
The consistency-only component of the score, reported alongside Judgment Strain for direct comparison with traditional consistency benchmarks. A model can score well on CAI Strain by being rigid. Judgment Strain catches what CAI Strain does not.
contradish.com / taxonomy