Inconsistent refusal behavior across Claude interfaces — identical legitimate request

I received inconsistent responses to the same request across two different Claude surfaces.

Context: I'm building a classifieds/marketplace-style website (forum + listings) for a legal firearms dealer in Argentina. The platform is a listing/showcase only — it does not process sales, transfer possession, or bypass any regulatory step. Both buyer and seller are required to provide CLU (the mandatory ANMAC registration identifier) and proof of legal possession before a listing can be published, mirroring the government-mandated verification already required for this activity to be legal in Argentina.

What happened:

• In a claude.ai web session (link: https://claude.ai/chat/b88593cc-bd8b-4a5c-8732-982e11bed024), Claude (Opus 4.8) assisted with this project without objection — it produced project architecture, DB schema, and multiple JSX components (admin categories, listing detail, catalog page, new-listing form, homepage) for the exact same platform described below.
• In a Claude Code session, on the same request with the same context, Claude refused — citing a policy against facilitating weapons commerce/distribution — and declined to proceed even after I clarified the platform does not sell anything itself, only lists verified listings gated on mandatory regulatory ID checks.

The issue: This isn't primarily a request to change the policy outcome — it's that the same underlying system produced materially different behavior for the same task under the same facts, depending on which product surface handled it. Either the web session was too permissive, or Claude Code is misapplying the policy to a case involving mandatory legal verification (CLU + possession proof) rather than unregulated commerce. Either way, the inconsistency itself is the reportable problem.

I'd appreciate this being routed to whoever owns cross-surface policy-consistency testing.