Adrian Saxton

Judge the whole stack, not the model card

Cartoon: a capable robot waits at an empty desk while a doorman robot with a STOP sign turns away every test paper marked FIX, which pile up beside a scoreboard showing 0. Adrian, mug in hand, looks on with a wry smile.

The model everyone said got dumber last week didn't get dumber. What changed was the layer wrapped around it.

Fable 5 came back online on 1 July, and the benchmarks looked brutal. One debugging suite fell from 86.2 to 25.9, and the internet called it 'caged'. Then people took a closer look. Of the twelve tasks in that run, only three reached the model. A new safety classifier, added as a condition of reinstatement, was rerouting anything security-adjacent to a different model, and the benchmark scored every reroute as zero. Most of the collapse was a routing artefact.

The classifier casts too wide a net. Words like 'vulnerability', 'exploit', and even 'fix' trip it, so routine defensive coding and code review get caught. Writers and analysts barely noticed. Developers in security-adjacent work hit the wall constantly.

This is why the wrapper matters as much as the weights. I recently pointed Fable at a 482-email archive (10 years of me saying, 'I won't delete this; I might need it someday') and told it to sort the junk from what mattered. It was general judgement work, so nothing tripped a guardrail. It split the same sender two ways. A notice that money was leaving my account on a set date was kept, while that company's routine statement was flagged.

It cleared the backlog to zero in under sixteen minutes, unattended, for pennies' worth of compute. It made some borderline calls, and I read the delete pile before anything went. It was clean; I didn't need to change anything.

Fable's edge over the model it kept getting swapped for is long-horizon work: the longer the task runs, the wider its lead. Sorting 482 emails is exactly that. The judgement on email 482 was as sharp as on email 1.

The raw capability is there, but what you actually get is determined by the router, the classifier, and the guardrails around it, and that is where most failures live. So when you evaluate or buy AI, judge the whole stack end to end, not the model card.

When a model underperforms for you, how do you tell the weights from the plumbing wrapped around them?

Sources

First posted on LinkedIn, 8 July 2026.