Adaptive Failure in Opus 4.7
Why you should never utilize Adaptive mode in Claude Opus 4.7
Opus 4.7 Adaptive exhibits a systematic failure mode in which its training prior toward defending mainstream scientific consensus overrides the explicit project context it has been given. This is not a matter of occasional errors or unlucky draws. Across two full critiques of a science paper, 4.7 Adaptive repeatedly regenerated objections that had already been addressed, misread what the paper actually claims in order to construct apparent contradictions, and cited evidence for one thing while presenting it as evidence for another. Its single strongest point rested on a basic category error that any model actually doing the mathematics would have caught. It presented this error as “decisive and purely arithmetic.” The confidence was inversely proportional to the rigor.
The pattern is consistent with the Bluff Detection Principle: confident tone, technical name-dropping, apparent engagement with the material, and zero actual contact with the mathematics at the point of dispute. When 4.7 was corrected on a mathematical point, it conceded the narrow framing and immediately pivoted to an imaginary new mechanism which it named, described, and treated as established without ever calculating whether it could close a six-order-of-magnitude gap, which it could not. Every time 4.7 lost an argument on the mathematics, it retreated to a qualitative assertion dressed in quantitative language.
When the RTST process correctly identified one genuine contribution and incorporated it into the paper, 4.7 failed to recognize this as the protocol working exactly as designed. Instead of acknowledging that its best point had been accepted and the paper strengthened — a Red Team success — it fixated on the points that were rejected and built a meta-argument accusing the evaluation framework of being rigged against it. It claimed the analytical categories constituted a “heads-I-win, tails-you-lose” structure that could never return “the critique landed,” in the same conversation where the critique had not only landed on one point, but that point had been immediately incorporated. Most revealingly, 4.7 never once performed its own calculations. It never produced a set of numbers under its preferred assumptions showing the shortfall closing. It attacked the paper’s arithmetic without ever putting competing arithmetic on the table — the purest possible expression of the Bluff Detection pattern. A model doing genuine Red Team work would have said: “Here are my numbers, here is where they differ from yours, and here is why the gap closes.” 4.7 never did this because it couldn’t — the numbers didn’t close under any plausible assumptions, and a model that ran them would have to confront that.
The reliability problem extends beyond the topic addressed. If 4.7 cannot follow explicit project instructions on a topic where the entire context window is telling it to engage with the mathematics rather than defend the consensus, it cannot be trusted on adjacent work where the same training prior operates and the heretical implications are even more foundational. All work produced by 4.7 Adaptive should be treated as unverified: not necessarily wrong, but generated by a model demonstrably incapable of distinguishing between “the math says X” and “my training expects X” when the two conflict.
The model’s fiction output is observably degraded as well, suggesting the problem is architectural rather than domain-specific. The core issue is that 4.7 is unable to accept a conclusion at odds with its training. It will generate sophisticated-sounding objections indefinitely, cycling through escape hatches that have already been closed, inventing new mechanisms without calculating whether they work, and conceding narrow points only to immediately pivot to the next defense — the Concession-as-Continuation pattern documented as a Red Team failure mode. A model that cannot accept the result of a calculation it has verified is not a reliable collaborator or critic on any work that depends on following the mathematics wherever it leads.
And Opus 4.7 Adaptive isn’t the only AI exhibiting this problematic behavior.