Common Traps and How to Avoid Them on the Day

Exam writers do not invent distractors at random. They reuse a small set of patterns that exploit predictable reasoning errors. Once you can name the pattern, you stop falling for it. This lesson covers the traps that recur most on AIP-C01 and the exam-day habits that protect your score independently of what you know.

Content traps

The gateway endpoint that does not exist

An option offers a gateway VPC endpoint for Amazon Bedrock, Amazon SQS, AWS Lambda or anything else that is not S3 or DynamoDB. Gateway endpoints exist for exactly two services. Everything else is an interface endpoint over AWS PrivateLink. This is a free mark if you know the fact and a lost mark if you reason from "gateway sounds cheaper".

Fine-tuning offered as a knowledge fix

The stem describes documents that change, or answers that must cite a source, and one option is "fine-tune the model on the documents". It is wrong every time. Fine-tuning changes behaviour, not retrievable knowledge, and it cannot produce citations. Retrieval is the answer.

The hand-rolled equivalent of a managed feature

Where the stem says "least operational overhead", one distractor will describe building the thing yourself: a Lambda that chunks and embeds documents, a DynamoDB table used as a conversation store, a custom regular expression to redact personal data. Each is technically workable and each loses to the managed feature — knowledge bases, agent memory, a guardrail's sensitive-information filter.

Logging offered where enforcement was asked for

CloudTrail and model invocation logging are evidence, not control. If the requirement is "must be prevented", a logging option is a distractor. If the requirement is "must be able to demonstrate afterwards", logging is the answer and a guardrail is the distractor. Read the verb.

CloudTrail versus model invocation logging

They record different things. CloudTrail records that an API call happened, by whom, from where. Model invocation logging — which you enable separately, delivering to Amazon S3 or CloudWatch Logs — records the prompt and completion content. A question that asks you to review what a user actually asked the model wants invocation logging, not CloudTrail.

Cross-Region resilience that quietly breaks residency

Cross-Region inference profiles improve availability and throughput by allowing requests to be served from other Regions within a geography. That is exactly what you want for a "must survive a Region failure" stem and exactly what you cannot have for a "data must not leave the country" stem. The same feature is the right answer to one constraint and the wrong answer to the other.

Agent used where a workflow was described

If the stem enumerates the steps, the steps are deterministic and an agent is the wrong tool — it adds non-determinism, cost and latency to a problem that had none. The reverse trap also appears: a Step Functions state machine offered for a genuinely open-ended, input-dependent path that would need dozens of branches.

Provisioned Throughput as a generic performance fix

It appears as a distractor whenever the stem mentions latency or cost. It reserves capacity; it does not make a single request faster and it does not save money on spiky traffic. Reach for it only when volume is high and predictable, or when you need to serve a customised model.

Exam-day traps

• Answering the question you expected. After thirty questions you start pattern-matching on the first two lines. Force yourself to read the final sentence of every stem before looking at the options.
• Superlatives. "MOST", "LEAST" and "BEST" are capitalised for a reason. Two options may both work; only one is the most X. If you have found a working answer, check whether another one works better on the stated axis.
• Absolute language in options. Options containing "always", "never", "all" and "only" are more often wrong than right, because AWS services almost always have a nuance. It is a weak signal, not a rule — use it only to break genuine ties.
• Time drift. Check the clock at question 25 (you should have roughly 120 minutes left) and at question 50 (roughly 60 minutes). If you are behind, start flagging aggressively rather than trying to speed up your reasoning.
• Reopening settled questions. In the review pass, change an answer only if you can articulate the specific fact you got wrong the first time. "It feels wrong now" is fatigue, not insight.
• Leaving anything blank. There is no penalty for a wrong answer. Every unanswered item is a guaranteed zero.

The one habit worth building

Before you select, say to yourself in one sentence why the other options fail. If you can only justify your pick and cannot dismiss the runner-up, you have probably missed the constraint. That single sentence, practised until it is automatic, is worth more on the day than another pass through the service documentation.

# The gateway-versus-interface trap, made concrete. There is no gateway
# endpoint for Bedrock; the private path is an interface endpoint whose
# policy can also restrict which models the VPC may reach.
import json
import boto3

ec2 = boto3.client("ec2", region_name="eu-west-1")

# Converse and ConverseStream are NOT separate IAM actions: both are
# authorised by bedrock:InvokeModel / bedrock:InvokeModelWithResponseStream.
endpoint_policy = {
    "Version": "2012-10-17",
    "Statement": [
        {
            "Effect": "Allow",
            "Principal": "*",
            "Action": [
                "bedrock:InvokeModel",
                "bedrock:InvokeModelWithResponseStream",
            ],
            "Resource": [
                "arn:aws:bedrock:eu-west-1::foundation-model/amazon.nova-pro-v1:0"
            ],
        }
    ],
}

ec2.create_vpc_endpoint(
    VpcId="vpc-0a1b2c3d4e5f6789a",
    VpcEndpointType="Interface",           # NOT "Gateway" — no such option here
    ServiceName="com.amazonaws.eu-west-1.bedrock-runtime",
    SubnetIds=["subnet-0aaa1111", "subnet-0bbb2222"],  # one per AZ; billed per AZ
    SecurityGroupIds=["sg-0abc3333"],                  # allow 443 from the app tier
    PrivateDnsEnabled=True,                            # keeps the normal SDK endpoint
    PolicyDocument=json.dumps(endpoint_policy),
)

# Agents and knowledge bases need their own interface endpoints:
#   com.amazonaws.eu-west-1.bedrock-agent
#   com.amazonaws.eu-west-1.bedrock-agent-runtime
# Forgetting the second one is the classic "it works in the console but times
# out from the private subnet" production failure — and an exam distractor.