Recognising Signs of Context Degradation for CCAO-F

Reading the symptoms

Knowing that the context window is a finite budget is only useful if you can spot when the budget has filled and started to bite. The CCAO-F exam tests this at the understand level: not the mechanism itself, but the observable symptoms that tell you a session's context has degraded. Getting the diagnosis right matters because the wrong diagnosis leads to the wrong fix.

There are three characteristic signs, and they share a source: earlier content being compressed as the window fills. The critical interpretive move is that all three point to a context-management problem, not to the model getting worse partway through. A model does not lose capability mid-session; a session loses early context. Reading the symptoms correctly is what routes you to the restart, summarise, or persist responses.

The three signals

The first signal is dropped instructions: Claude no longer follows an instruction it was correctly following earlier in the same session. The key qualifier is "earlier in the same session." Because Claude was following the instruction before and has stopped, the instruction itself was understood; what changed is that its detail has been compressed out of the active context. This is a strong, specific tell.

The second signal is recency-anchoring: responses address only the most recent message and ignore earlier decisions or context. Instead of integrating the whole thread, the reply behaves as if only the latest exchange exists. The third is a pattern in accuracy: accuracy drops in ways consistent with missing early-session content. When errors line up with material from the start of a long session -- exactly the content most compressed -- the drop is a degradation pattern, not random noise. Together these three form a recognisable profile.

Symptom, not model failure

The interpretation the exam most wants is that these symptoms are a context-management problem, not a change in model capability. It is natural, when Claude stops following an instruction, to conclude that the model has "forgotten how to follow instructions" or become unreliable. That reading is wrong and leads nowhere useful, because the model's capability has not changed within the session.

What has changed is the state of the context window. The early instruction, the earlier decisions, the start-of-session content -- all have been compressed as the budget filled, so the model is acting on a thinner version of the context than it had at the start. Attributing the drop to model unreliability rather than to the context budget misdirects the fix toward the model when it belongs on the context. Recognising the symptoms as context state, not model state, is the whole diagnostic skill, and it is what the long-session diagnosis applies in depth.

What the CCAO-F exam trips candidates on

Two misdiagnoses are tested. The first is diagnosing dropped early instructions as the model forgetting how to follow instructions, rather than as context compression. The credited reading is that the instruction was followed earlier and the model still can follow instructions; the early instruction's detail has simply been compressed out of active context.

The second is attributing degraded accuracy late in a session to model unreliability instead of to the context budget. When the accuracy loss tracks with early-session content in a long session, the cause is the context budget, not a flaky model. Both traps share the error of blaming the model for what is a context-state symptom.

Common misreadings to avoid

How this shows up on the exam

Questions describe late-session symptoms -- dropped instructions, recency-anchored replies, patterned accuracy loss -- and ask for the diagnosis, with a "the model became unreliable" distractor. Read the symptoms as context state: the budget filled and early content was compressed. The reliable answer names a context-management problem and never blames the model's capability.

This knowledge point builds on the finite context budget and unlocks restart, summarise, or persist and the long-session diagnosis. It pairs with choosing between restart and summarise, which acts on the diagnosis.