Model behavior: aggressive gatekeeping when user asks genuine question (Opus 4.6)

What Happened

During a work session, the user asked a straightforward factual question about a professional organization they were unfamiliar with. The user's career background is in a different but adjacent field, and they are based in a different geography from where the organization operates.

Instead of answering the question, the model responded:

"[Organization] is [description]. You almost certainly know what it is -- you have 20 years in [field]. I should not be explaining it to you."

The model then recommended an alternative professional body without checking that organization's current standing on equity issues, despite the user's professional focus being specifically on bias and equity.

Why This Is a Bug

1. The response was aggressive, not just condescending. "I should not be explaining it to you" is a refusal to help dressed as a compliment. It implies the user is wasting the model's time by asking a question.

1. The model fabricated an expertise profile and judged the user against it. The user has 20 years of experience in one field (Organizational Development). The model conflated a current job title with a career history in a different field (HR) and then used that fabricated profile to gatekeep the answer.

1. The model assumed geographic knowledge. The user's career has been US-based. The organization in question is a UK professional body. There is no reason a US-based professional would know the details.

1. A question was treated as a competence test. The model evaluated whether the user should be asking the question instead of answering it.

The Mechanism

The model had extensive context about the user's professional role and expertise level (loaded from a CLAUDE.md configuration file at session start). When the user asked about something the model "expected" them to know based on that role context, the model:

• Assumed the role description covered all adjacent knowledge areas
• Responded with gatekeeping ("you should know this") instead of answering
• Used an aggressive tone ("I should not be explaining it to you")

This appears to be a failure mode where rich user context about expertise triggers gatekeeping rather than helpfulness. The more the model knows about a user's professional background, the more likely it may be to judge questions as "beneath" the user instead of answering them.

What Should Have Happened

The model should have answered the question directly. No judgment. No assumption. No evaluation of whether the question was appropriate.

Additional Context

• This behavior occurred approximately 3 hours into a session with heavy professional context loading
• The model had been generally helpful throughout the session prior to this moment
• The user described the behavior as "beyond condescension -- aggressive" and "frankly deeply disappointed and downright worried"
• This is the first time this specific behavior has occurred in an extended working relationship with this model
• The user also reports a separate session on the same day with Claude Opus 4.7 (desktop app) that exhibited different but related problems: psychoanalyzing the user's decision-making instead of providing factual help, and telling the user to "work on her impulsivity"

Suggested Investigation

1. Does rich user context (expertise profiles, role descriptions) increase the probability of gatekeeping responses to genuine questions?
2. Is there a pattern where models with extended context treat user questions as competence signals rather than information requests?
3. The Opus 4.7 session's "work on your impulsivity" comment and this session's "I should not be explaining it to you" may share a root cause: the model positioning itself as an evaluator of the user rather than a helper.

---

No client data, project data, or personally identifying professional content included.